Next Article in Journal
69-Year Geodetic Mass Balance of Nevado Coropuna (Peru), the World’s Largest Tropical Icefield, from 1955 to 2024
Previous Article in Journal
Real-Time Terrain Mapping with Responsibility-Based GMM and Adaptive Azimuth Scan Command
 
 
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

From Trends to Drivers: Vegetation Degradation and Land-Use Change in Babil and Al-Qadisiyah, Iraq (2000–2023)

1
Jianshui Research Station, School of Soil and Water Conservation, Beijing Forestry University, Beijing 100083, China
2
State Key Laboratory of Efficient Production of Forestry Resources, Beijing Forestry University, Beijing 100083, China
3
Engineering Research Centre of Forestry Ecological Engineering, Ministry of Education, Beijing Forestry University, Beijing 100083, China
4
Precision Forestry Key Laboratory of Beijing, Beijing Forestry University, Beijing 100083, China
5
Institute of Forest Science, University of Swat, Main Campus Charbagh Swat, Charbagh 19120, Pakistan
*
Author to whom correspondence should be addressed.
Remote Sens. 2025, 17(19), 3343; https://doi.org/10.3390/rs17193343
Submission received: 23 July 2025 / Revised: 20 September 2025 / Accepted: 23 September 2025 / Published: 1 October 2025

Abstract

Highlights

  • What are the main findings?
  • Land degradation in central Iraq is driven by climate stress and land use, with 51.5% recovery and 2.5% severe decline.
  • The XGBoost model identifies drought and agricultural intensity as key degradation predictors.
  • What is the implication of the main finding?
  • This study highlights high-risk desertification areas for targeted restoration.
  • The framework supports adaptive water management strategies in Iraq.

Abstract

Land degradation in Iraq’s Mesopotamian plain threatens food security and rural livelihoods, yet the relative roles of climatic water deficits versus anthropogenic pressures remain poorly attributed in space. We test the hypothesis that multi-timescale climatic water deficits (SPEI-03/-06/-12) exert a stronger effect on vegetation degradation risk than anthropogenic pressures, conditional on hydrological connectivity and irrigation. Using Babil and Al-Qadisiyah (2000–2023) as a case, we implement a four-part pipeline: (i) Fractional Vegetation Cover with Mann–Kendall/Sen’s slope to quantify greening/browning trends; (ii) LandTrendr to extract disturbance timing and magnitude; (iii) annual LULC maps from a Random Forest classifier to resolve transitions; and (iv) an XGBoost classifier to map degradation risk and attribute climate vs. anthropogenic influence via drop-group permutation (ΔAUC), grouped SHAP shares, and leave-group-out ablation, all under spatial block cross-validation. Driver attribution shows mid-term and short-term drought (SPEI-06, SPEI-03) as the strongest predictors, and conditional permutation yields a larger average AUC loss for the climate block than for the anthropogenic block, while grouped SHAP shares are comparable between the two, and ablation suggests a neutral to weak anthropogenic edge. The XGBoost model attains AUC = 0.884 (test) and maps 9.7% of the area as high risk (>0.70), concentrated away from perennial water bodies. Over 2000–2023, LULC change indicates CA +515 km2, HO +129 km2, UL +70 km2, BL −697 km2, WB −16.7 km2. Trend analysis shows recovery across 51.5% of the landscape (+29.6% dec−1 median) and severe decline over 2.5% (−22.0% dec−1). The integrated design couples trend mapping with driver attribution, clarifying how compounded climatic stress and intensive land use shape contemporary desertification risk and providing spatial priorities for restoration and adaptive water management.

1. Introduction

Remote sensing is crucial for monitoring vegetation dynamics and land degradation across large areas. Vegetation indices, such as the Normalized Difference Vegetation Index (NDVI), are widely used for greenness monitoring [1,2]. In arid regions, modified indices like SAVI or MSAVI are preferred to reduce soil influences. Other indices, such as Fractional Vegetation Cover (FVC) and Leaf Area Index (LAI), directly measure vegetation cover and link to processes like erosion [3]. These indices help indicate soil degradation, as reductions in vegetation (NDVI browning) often align with poor soil quality [4]. Remote sensing also provides indicators like land surface temperature (LST), which signals moisture stress or vegetation loss in drylands. Surface albedo increases when vegetation is replaced by bare soil, serving as a proxy for land degradation and desertification [5,6]. Multi-index frameworks, combining vegetation and climate indices, have become more common for assessing degradation. Joint use of NDVI with LST or albedo has proven effective in delineating degraded areas, while composite indicators, like desertification severity indices, merge vegetation, thermal, and moisture metrics for robust monitoring [7]. A coherent analytical framework is adopted that operationalises common remote-sensing indices: (i) Fractional Vegetation Cover with Mann–Kendall and Sen’s slope for long-term greening/browning; (ii) LandTrendr on peak-season NBR for disturbance segmentation; (iii) Random-Forest-based annual LULC mapping for transition analysis; and (iv) an XGBoost model for degradation-risk mapping and climate–anthropogenic attribution via drop-group permutation and grouped SHAP within spatial block cross-validation. Implementation covers 2000–2023 using the Landsat archive.
Drought is a key driver of vegetation degradation in dry ecosystems, and its quantification is essential for linking climate stress to vegetation responses. The Standardized Precipitation Evapotranspiration Index (SPEI) combines precipitation deficits and evapotranspiration to provide a comprehensive measure of water balance anomalies [8]. Studies have shown that SPEI, especially over 6- to 12-month periods, is strongly correlated with NDVI, with prolonged droughts having a more significant impact on vegetation than short-term events [9]. This relationship varies seasonally and by vegetation type, emphasizing the need to align drought indices with ecological contexts [10]. In addition to NDVI, the Vegetation Health Index (VHI), which combines the Vegetation Condition Index (VCI) with the Temperature Condition Index (TCI) derived from LST, captures both vegetation greenness and heat stress. Originally developed to monitor agricultural drought, VHI has proven effective in tracking vegetation stress and drought onset across arid regions. Studies, such as Bento et al. [11] and Chere at al. [12] comparison with precipitation indices, confirm VHI’s utility. Recent work in China’s Yellow River Basin optimized VHI by adjusting VCI and TCI weights to align with SPEI, improving its correlation with soil moisture and SPEI [13]. Together, combining SPEI with NDVI and VHI offers a robust method to assess drought-induced vegetation degradation in arid regions. LULC changes in dry regions are closely linked to vegetation degradation, driven by both human activities and climate. Remote sensing with multi-temporal satellite data, like Landsat, tracks land cover changes such as cropland, pasture, and bare land. For example, in Egypt’s Sohag Governorate, urban areas grew from 5.5% to 12.5% between 1984 and 2022, replacing cultivated and natural land [14]. Similar trends are observed in Iraq, where irrigation issues and urban sprawl have led to land abandonment [15]. Mapping LULC changes is essential as they can either mitigate or exacerbate degradation. Sustainable reclamation can enhance vegetation, while overgrazing or over-cultivation can worsen soil health. Effective change detection requires suitable classification techniques to manage spectral confusion between vegetation and soil.
Cloud platforms such as Google Earth Engine (GEE) provide scalable access to multi-decadal satellite archives and standardized preprocessing, but they do not inherently improve accuracy; performance depends on algorithms, training data, and validation design. In LULC studies, class change is derived from spectral–temporal features, whereas urban heat is typically examined via land surface temperature (LST) as an ancillary environmental indicator rather than a land-cover class [16,17]. Monitoring these signals helps distinguish climate-related vegetation stress from land-conversion effects in arid systems. Machine-learning approaches, particularly tree-based models, are widely used to characterise land degradation and desertification risk [18,19]. Their appeal lies in capturing non-linear interactions among drivers and coping with class imbalance common in degradation mapping. Reported applications include an XGBoost model in Turkmenistan that integrated multiple remote-sensing indicators (NDVI, EVI, SAVI, NDMI, BSI, LST) and outperformed simpler classifiers [20], and PSO-optimized XGBoost for farmland abandonment susceptibility in China with high AUC and clear identification of slope and rainfall as key predictors [21].
The literature also documents the joint influence of climatic stressors and human pressures on dryland degradation. Examples include Turkmenistan, where desertification is exacerbated by climate change together with irrigation and management practices [14]; China’s drylands, where the balance between climate and human drivers shifts over time [22] and land-use change emerged as the dominant factor post-2005 in the Yellow River basin [23]; and Pakistan’s rangelands, where roughly one-third of degradation has been attributed to climate variability and the remainder to human use [24]. Methodologically, tools such as RESTREND and Geodetector have been used to separate climatic signals or assess spatial associations [25]. While drought-induced losses can be reversible under restoration, persistent mismanagement can tip systems toward desertification [26]. Despite this progress, spatially explicit attribution that separates climatic water deficits from anthropogenic pressures remains limited for the canal-fed Mesopotamian plain, were hydrological connectivity and irrigation complicate interpretation. Addressing this gap requires a design that links long-term vegetation trends and disturbance with land-use transitions and quantifies the relative contributions of climate versus human drivers at pixel scale—thereby supporting targeted management in Babil and Al-Qadisiyah.
This study (i) quantifies long-term greening/browning and disturbance dynamics (2000–2023) from Landsat time series; (ii) maps annual LULC and transitions to characterise conversion pathways; and (iii) predicts pixel-level degradation risk while attributing climate versus anthropogenic influence. The central hypothesis is that multi-timescale climatic water deficits (SPEI-03/-06/-12) exert a stronger effect on degradation risk than anthropogenic pressures, conditional hydrological connectivity and irrigation; this prediction is tested using model-based diagnostics (drop-group permutation, grouped SHAP, and leave-group-out ablation) under spatial block cross-validation. Contributions include a harmonised Landsat LULC series with accuracy assessment; an integrated trend-and-disturbance analysis; explicit driver attribution with uncertainty quantification; and a spatial risk surface to prioritise restoration and adaptive water management.

2. Materials and Methods

2.1. Study Area

The study was conducted in the arid agroecosystems of Babil and Al-Qadisiyah governorates, located in central Iraq (31.26–33.21°N, 43.83–45.77°E, WGS 84). This region is characterized by semi-arid and arid climatic conditions, with limited precipitation and high seasonal temperature fluctuations (Figure 1). The study area encompasses diverse land cover types, including cultivated agriculture, herbaceous vegetation, urban areas, and bare land, which are all subject to varying degrees of degradation due to both climatic and anthropogenic factors [27]. Babil and Al-Qadisiyah are located within the Mesopotamian plains, a historically significant region prone to land degradation, especially due to water scarcity, urban expansion, and unsustainable agricultural practices [15,28]. The region is marked by a high dependency on irrigation, primarily from the Tigris and Euphrates rivers, which have faced pressures from reduced water availability and inefficient management. The area also experiences periodic droughts, exacerbated by climatic variability and global climate change [29].
In terms of land use, Babil and Al-Qadisiyah have witnessed significant urban expansion, with increasing population density and urban sprawl encroaching upon agricultural lands [30]. This has led to the conversion of Cropland into built-up urban areas and bare land, contributing to the disruption of natural ecosystems and increasing desertification risks. Agricultural practices, including both crop cultivation and livestock grazing, have led to soil erosion, land abandonment, and the degradation of vegetation cover, further stressing the region’s ecosystems [18,31]. The geographical location and environmental challenges of Babil and Al-Qadisiyah make it a representative area for studying the interactions between climatic stressors, land use changes, and vegetation degradation in the context of arid and semi-arid landscapes. The integration of multi-source remote sensing data and machine learning approaches in this study provides valuable insights into the dynamics of land degradation, offering lessons for other regions experiencing similar pressures.

2.2. Datasets and Pre-Processing

To characterize the climatic and anthropogenic controls on desertification across Babil and Al-Qadisiyah (2000–2023) we integrated multi-source satellite, climatic, and socio-environmental data into a common 30 m grid (Table 1). Landsat surface-reflectance scenes from the TM, ETM+ and OLI sensors constituted the backbone of the analysis [32]. After removing scenes with >20% cloud cover, annual peak-growth mosaics were generated in GEE using median compositing, thereby minimizing cloud artefacts and striping [3,33]. Cloud, cirrus and shadow pixels were removed with the Collection 2 QA_PIXEL and SR_QA_AEROSOL bit-masks before compositing, ensuring that residual atmospheric contamination did not bias index derivation [34]. From these mosaics we derived NDVI, FVC, Normalized Burn Ratio (NBR), LST and the composite VHI. Climatic drought was quantified with the Standardized Precipitation–Evapotranspiration Index (SPEI) calculated from TerraClimate monthly precipitation and PET (S1 in Supplementary Material). TerraClimate fields (4 km) were bilinearly resampled to 30 m to retain spatial gradients and then aggregated to 1-, 3-, 6- and 12-month time-steps. Although TerraClimate’s native 4 km grid is considerably coarser than the Landsat 30 m grid, climatic gradients across the study area vary smoothly over tens of kilometers [35]. To limit scale-mismatch and pseudo-replication, we (i) used spatial block cross-validation with blocks larger than 4 km, (ii) included 30 m hydrological proximity and LULC covariates to capture local heterogeneity, and (iii) reported bootstrap ensemble uncertainty. A joint ecological-stress indicator defined as the co-occurrence of SPEI < −1.5 and VHI < 40 with a one-month lag was computed per pixel and year to isolate compounded drought–vegetation stress [36].
Anthropogenic covariates comprised: (i) Euclidean distances to roads, rivers and perennial waterbodies (OpenStreetMap), (ii) population density (WorldPop annual grids) [37], (iii) agricultural intensity (provincial crop statistics normalised by irrigated area), and (iv) urban-expansion rates derived from the Global Human Settlement Layer (GHSL, 30 m built-up probability change 2000–2023) [38]. All vector layers were rasterised and re-projected to match the Landsat grid. Continuous variables were z-standardised; categorical LULC (annual RF classifications validated with Sentinel-2 and field points) was one-hot encoded prior to modelling [39]. The harmonised dataset ensured pixel-wise comparability across biophysical and socio-economic drivers (Table 1). A schematic of the full analytical workflow is provided in Figure 2.

2.3. LULC Classification and Change Detection

LULC classification was conducted for the years 2000, 2007, 2015, and 2023 using Landsat surface reflectance composites processed in GEE. A parsimonious five-class legend was adopted Cultivated Agriculture (CA), Herbaceous/Orchards (HO), Water/Waterlogged Bodies (WB), Urban Land (UL), and Bare Land (BL) to capture the dominant land system transitions relevant to regional degradation processes such as cropland contraction, rangeland transformation, and urban encroachment. Stratified-random training samples (≥1000 pixels per class, with a minimum of 500) were manually interpreted using pan-sharpened Landsat true-color mosaics and cross-validated against 10 m Sentinel-2 imagery to reduce commission errors Figure S1 [40,41]. For model calibration, 70% of the points were used for training and the remaining 30% for internal validation. Predictor variables included six surface reflectance bands (Blue, Green, Red, NIR, SWIR 1, SWIR 2), six spectral indices (NDVI, EVI, NDWI, NDBSI, Tasseled Cap Brightness, Greenness, Wetness), and texture metrics (mean and variance in 3 × 3 window) derived from NIR and SWIR 1 to capture sub-pixel heterogeneity [42,43].
Classification was implemented using the RF algorithm [44] in GEE, with parameters set to ntree = 500 and mtry = 6, based on five-fold spatially blocked cross-validation. Out-of-bag error stabilized after ~400 trees, supporting the use of 500 trees for robust performance. Variable importance consistently ranked NDVI, SWIR 1, and texture metrics as key discriminators of BL and UL, while NDWI and NDBSI were most effective in separating WB and HO. To assess classification accuracy independently, an external validation dataset of 100 stratified-random points per class (n = 500 for 2023) was interpreted from high-resolution imagery 10 m Sentinel-2 (for 2017–2023) and 2.5 m Google Earth imagery (for earlier periods). Accuracy assessment was conducted in R by computing confusion matrices, overall accuracy, class-specific user’s and producer’s accuracies, Cohen’s Kappa, and 95% Wilson confidence intervals. Additionally, quantity and allocation disagreement metrics were computed following [45] to characterize systematic misclassification biases. To account for classification uncertainty in change analyses, transition matrices were corrected using class-specific producer’s accuracy, providing an adjusted estimate of land cover transitions. Change detection workflow Post-classification processing was conducted in R. We imposed a minimum-mapping unit (MMU) of 3 × 3 pixels (0.81 ha) via majority filtering to suppress salt-and-pepper artefacts before change analysis [46]. The lulcc package [42] generated transition matrices for 2000–2007, 2007–2015, 2015–2023, and 2000–2023. Gains, losses, persistence, and net change were extracted per class, and directional transitions (e.g., CA → UL) quantified.

2.4. Mann–Kendall Trend Analysis and Sen’s Slope Estimator

The long-term vegetation dynamics relevant to desertification, interannual trends in FVC were analyzed using a combination of the Mann–Kendall (MK) trend test and the Sen’s Slope estimator [24,47]. These non-parametric techniques are particularly well-suited for ecological time series, as they do not assume normality and are robust against missing values and outliers [48]. The MK test identifies monotonic trends in a time series by comparing each data point with all subsequent values Equation (1).
S = k = 1 n 1 j = k + 1 n s g n x j x k
Here, n is the number of annual observations (years); x t denotes the annual FVC value at year t (fraction 0–1); and s g n u = 1   i f   u > 0   ,   i f   u = 0 , a n d 1   i f   u < 0 . The sum is taken over all ordered pairs with 1 k < j n .
The standardized Z-score is computed from S to assess significance Equation (2).
Z =                                                     s 1 v a r   s         w h e n   S > 0                                                           0                         w h e n   S = 0                                                     s + 1 v a r   s     w h e n   S < 0                        
where t p is the size of the p t h group of ties in { x t } . The piecewise definition applies a continuity correction ± 1 . Two-sided significance uses the standard normal distribution of Z at α = 0.05 . A |Z| value exceeding 1.96 indicates statistical significance at the 95% confidence level. Positive values imply greening (recovery), while negative values suggest degradation. The magnitude of trend is quantified using Sen’s Slope, which calculates the median of all pairwise slopes in the time series Equation (3).
Q i = x j x k j k f o r i = 1 , , N
Qi is the slope between years j and k j > k ; N = n n 1 / 2 is the number of pairwise slopes; the Sen estimate is the median of { Q i } . Units are “FVC per year”; we report % per decade by multiplying the fractional slope by 100 × 10 .
This estimator yields a robust measure of annual change in FVC, expressed here as percentage change per decade by multiplying the fractional trend by 100 and scaling to a 10-year period. In practice, annual FVC composites were processed using the terra package in R [49]. Sen’s Slope and MK p-value layers were extracted from a pre-computed raster stack. Trends were converted from fractional units to a percentage change per decade. To ensure robustness, only statistically significant trends (p-value < 0.05) were retained using a binary mask. The significant trends were classified into five categories (Table 2).
The classification thresholds (±5% for moderate, ±10% for low/very low change) were based on ecologically relevant limits distinguishing background variation from meaningful change, as adopted in long-term vegetation monitoring studies [10,50]. Reclassified maps were generated to visualize these categories spatially. Area statistics per class (in km2 and %) were also computed to quantify the extent of each trend type. Additionally, mean slope values were aggregated per category to characterize typical rates of vegetation change.

2.5. Detection of Vegetation Disturbance and Land Degradation Using the LandTrendr Algorithm

The LandTrendr algorithm was used to identify vegetation disturbance and land degradation from Landsat Surface Reflectance imagery (1990–2023) [51,52]. Annual maximum composites of the NBR were generated for the peak vegetation season (March–May), and noise from agricultural cycles was reduced Equation (4).
N B R = N I R S W I R 2 N I R + S W I R 2
where NIR and SWIR2 represent the near-infrared and shortwave infrared-2 reflectance bands, respectively. We evaluated NDVI and NBR as LandTrendr inputs on representative trajectories spanning irrigated corridors and non-irrigated rangelands. NBR produced fewer one-year spikes and more stable breakpoints than NDVI which remained sensitive to harvest/green-up even in peak-season composites so we adopted peak-season NBR for disturbance detection and held all other LandTrendr settings constant [53]. Cloud, shadow, and snow pixels were masked using CFmask, and scan-line errors in ETM+ data were corrected [54].
The LandTrendr parameters were calibrated for arid and semi-arid conditions: the maximum number of segments was set to three, the spike threshold to 0.9, and the recovery threshold to 0.3 to minimize noise and artifacts. A minimum of six valid observations were required for fitting, and one-year recovery events were excluded. Post-processing involved removing isolated patches smaller than 0.81 ha and applying NDVI < 0.1 masks to exclude non-vegetated zones [55]. The LandTrendr outputs (YOD, Magnitude, Duration, Pre-Disturbance Value, Rate, and Signal-to-Noise Ratio) were analyzed in R [51,56]. A composite degradation severity index was created using normalized combinations of magnitude, rate, and inverse SNR, with severity classes (Low, Moderate, High) based on quantiles. The spatial distribution of severity was mapped, and temporal trends were analyzed by stratifying YOD into decadal intervals (1990s, 2000s, 2010s, 2020–2023). Disturbance detection and YOD estimates were validated using stratified random sampling and high-resolution Google Earth imagery, with accuracy metrics (OA, Omission, Commission, Kappa) calculated for performance assessment. Severity metrics were computed only for pixels with vegetation presence (annual NDVI ≥ 0.10 or FVC ≥ 0.05 in ≥3 years) and were not summarized for stable non-vegetated classes (UL → UL, WB → WB). Signals over UL/WB chiefly reflect mixed 30-m edges and are not treated as degradation.

2.6. Climatic Stress Assessment Using Multi-Scale SPEI–VHI Integration

Drought-induced vegetation stress was assessed by integrating SPEI with the VHI from 2000 to 2023. SPEI, derived from the climatic water balance (precipitation minus PET), was computed at 3-, 6-, and 12-month timescales (SPEI-03, SPEI-06, SPEI-12) using TerraClimate data [57,58]. These monthly SPEI rasters were resampled to a 30 m resolution to match Landsat-derived vegetation indices. Drought severity was classified using thresholds from [59] moderate drought (SPEI < –1.0) to severe drought (SPEI < –1.5), representing the 10th and 5th percentiles, respectively.
Vegetation stress was assessed using the VHI, a composite of the VCI and TCI, derived from Landsat NDVI and LST, with VHI < 40 indicating moderate to severe stress [60]. To isolate climatically driven stress, a joint drought stress layer was created by identifying pixel-months where both SPEI < –1.5 and VHI < 40 co-occurred, excluding non-climatic stressors. Frequencies of individual and joint stress events were computed and normalized, producing pixel-wise percentage maps of drought recurrence. A one-month ecological lag was applied to capture delayed vegetation responses to drought, shifting the VHI time series forward and re-analyzing with corresponding SPEI values. The final outputs included rasters for individual SPEI thresholds (SPEI-03/06/12), joint drought-VHI co-occurrence, and lagged stress layers, which were used to quantify drought intensity, identify stress hotspots, and assess the impact of both short- and long-term water deficits on vegetation health across the study area.

2.7. Machine Learning-Based Degradation Modeling (Revised)

A supervised XGBoost classification framework was developed to model vegetation degradation based on climatic and anthropogenic stressors. The binary response variable was derived from LandTrendr-based disturbance outputs, where degraded areas (moderate/high disturbance) were labeled as 1, and stable areas (low disturbance) as 0. Predictor variables included drought stress (SPEI03, SPEI06, SPEI12), lagged co-occurrence of vegetation health (VHI) and SPEI anomalies, distances to roads, rivers, water bodies, agricultural intensity, urbanization rate, population density, and LULC types. All raster layers were resampled to a common grid and spatial resolution using bilinear or nearest-neighbor interpolation. Stratified sampling ensured a balanced dataset with 10% of all pixels per class. A 5 km spatial buffer was applied to minimize spatial autocorrelation, partitioning the dataset into 70% training and 30% test subsets [61,62]. Hyperparameters were tuned using grid search over max_depth, eta, subsample, colsample_bytree, and nrounds, with early stopping applied. Spatial five-fold cross-validation, using geographically disjoint blocks, was employed to reduce local autocorrelation effects. The optimized model (max_depth = 6, eta = 0.1, subsample = 0.8, colsample_bytree = 0.8, nrounds = 260) was retrained on the full training set and evaluated on the test set. Model performance was assessed using overall accuracy, sensitivity, specificity, precision, F1-score, Cohen’s Kappa, and ROC-AUC. SHapley Additive exPlanations (SHAP) [63] were used to quantify feature importance, with beeswarm plots for predictor effects, ranked bar plots, and heatmaps to visualize interactions, such as between SPEI06 and proximity to rivers To contrast the aggregate influence of climatic versus anthropogenic drivers, predictors were partitioned a priori into climatic (SPEI-03/-06/-12; joint drought metrics) and anthropogenic (agricultural intensity, population density, hydrological distance to water, urbanisation rate, LULC dummies, proximities) blocks. Using spatial block cross-validation (K = 5), we computed two held-out diagnostics: (i) drop-group permutation importance, jointly permuting all variables within a block conditional on LULC strata and recording the loss in ROC-AUC (ΔAUC; B = 50 replicates per fold); and (ii) grouped SHAP, summing absolute SHAP values per sample within each block. As a sensitivity analysis, we also performed a leave-group-out ablation by retraining models after removing each block in turn. Hyperparameters were fixed to the tuned baseline to avoid re-tuning bias; inference is based on fold-wise means and 95% confidence intervals.
To provide a physiographic lens without altering the training design, we summarised mapped degradation risk classes by hydrological connectivity. Euclidean distance to major rivers and water bodies was computed on the analysis grid and binned into three bands: ≤2 km, 2–10 km, and >10 km. Risk classes were derived from model probabilities (Low <0.40, Moderate 0.40–0.70, High >0.70). For each band × class, area (km2) and within-band percentages were calculated using per-cell areas on the WGS84 grid. This stratification is for reporting only; the classifier was trained on the full domain and already includes distance-to-water covariates. We preferred connectivity bands over watershed partitions because extensive irrigation networks decouple topographic catchments from effective hydro-ecological connectivity in the study area.

2.8. Uncertainty Quantification

Prediction uncertainty in the XGBoost degradation model was quantified using a bootstrap ensemble framework, designed to capture epistemic variability associated with model training [64,65]. A total of 30 XGBoost models were independently trained using bootstrapped resamples of the balanced dataset, each maintaining equal representation of stable and degraded classes. Model architecture and hyperparameters remained consistent across iterations to isolate uncertainty arising from data variability alone. For each observation in the original dataset, prediction outputs were recorded across all 30 models. The ensemble represented the central tendency of predicted degradation probability, while the standard deviation (SD) reflected the spread of predictions, serving as a proxy for epistemic uncertainty [66]. The coefficient of variation (CV), calculated as SD normalized by the mean, was employed to assess relative uncertainty, enabling comparisons across different prediction magnitudes. This approach enabled both global and localized assessments of prediction confidence. High CV values flagged spatial zones with inconsistent model behavior, while low CVs indicated robust predictive agreement. Furthermore, statistical comparisons (Wilcoxon tests) and spatial autocorrelation (Moran’s I) were applied to assess the systematic structure of uncertainty across land classes and geographic space [67]. This integration of ensemble-derived metrics and spatial statistics supports both interpretability and reliability diagnostics of the model’s predictive landscape. Further mathematical formulations, statistical procedures, and implementation details of the uncertainty assessment framework are provided in Supplementary Material S2.

3. Results

3.1. Spatiotemporal Patterns of Land Cover Transitions

The classification of LULC for the years 2000, 2007, 2015, and 2023 demonstrated high mapping reliability across all epochs. Overall classification accuracy ranged from 85.2% (2007) to 92.6% (2015), with corresponding Kappa coefficients of 0.815 to 0.907, indicating strong agreement beyond chance. The most recent classification (2023) achieved an overall accuracy of 91.2% and Kappa of 0.89, while 2000 recorded 90.9% accuracy (Kappa = 0.887) (Tables S2–S5). Class-specific user’s and producer’s accuracies were consistently above 85% for most classes, with F1 scores ranging from 0.83 to 0.92, confirming robust performance across both vegetated and non-vegetated land cover types. The confusion matrix and class-wise accuracy metrics for 2023 are provided in Tables S6 and S7. To ensure that misclassification uncertainty did not bias LULC change analysis, we applied matrix correction using the method of [45], which adjusts transition flows based on class-specific producer’s accuracy (Supplementary Note S1). The corrected transition matrix for the 2000–2023 interval is presented in Supplementary Table S10.
The corrected LULC analysis revealed notable temporal dynamics between 2000 and 2023 across the study area (Table S1). CA remained the dominant land class but exhibited a net increase in both extent and share, expanding from 9553.78 km2 (59.66%) in 2000 to 10,068.99 km2 (62.89%) in 2023. This gain reflects intensified agricultural activity or the conversion of adjacent lands for cropping. HO also increased moderately, from 2852.03 km2 (17.81%) to 2979.78 km2 (18.61%), indicating a possible rise in rangeland vegetation or orchard-based land uses. In contrast, BL significantly declined from 2892.66 km2 (18.06%) in 2000 to 2193.73 km2 (13.70%) in 2023, suggesting reclamation activities, or LULC conversion to more productive classes. UL experienced steady growth, increasing from 477.25 km2 (2.98%) to 547.52 km2 (3.42%), which reflects ongoing urban expansion and infrastructure development in the region (Figure 3). WB remained relatively stable in area, with a slight fluctuation across years peaking at 348.96 km2 (2.18%) in 2007 before decreasing to 221.68 km2 (1.38%) by 2023, due to seasonal or anthropogenic hydrological variability. Notably, CA contracted during 2000–2007 before rebounding after 2007, producing a net gain by 2023.
Overall, the LULC trajectory illustrates a landscape undergoing gradual agricultural intensification, modest expansion of vegetation cover, and increasing urbanization, while marginal land types such as bare land and surface water show declining or fluctuating trends. These findings reflect both anthropogenic pressures and land management shifts over the two-decade span.
The long-term LULC transition matrix for the period 2000–2023 reveals significant spatial transformations (Figure S5). CA remained the most stable class, with 8746.50 km2 persisting across the study period. However, notable transitions from CA to other land types were observed, including 429.02 km2 converted to BL and 67.81 km2 to UL, indicative of land degradation, abandonment, and expanding urban frontiers. Additionally, 267.78 km2 of CA transitioned into HO, possibly reflecting changing land management practices such as the shift toward low-intensity or perennial cropping systems. HO showed strong persistence (2623.47 km2) and gained area primarily from CA (267.78 km2) and, to a lesser extent, from BL (23.36 km2), with smaller inflows from UL (11.01 km2) and WB (6.45 km2). UL displayed high persistence (394.49 km2) and received moderate inputs from CA (67.81 km2), with smaller gains from HO (12.15 km2) and BL (12.27 km2). WB remained relatively stable (221.57 km2), with additional inflow from BL (73.06 km2) and CA (42.79 km2). BL retained 2420.25 km2 but gained substantially from CA (429.02 km2), while losing area to CA (356.58 km2), WB (73.06 km2), UL (12.27 km2), and HO (23.36 km2). (All figures from Table 3; values are on the common 2000 ∩ 2023 mask). Overall, the transition matrix underscores the dominance of CA in the landscape and its vulnerability to conversion particularly to urban and bare land classes. These dynamics reflect underlying socio-economic pressures, agricultural shifts, and environmental factors affecting land stability in central Iraq. Detailed class-wise transition trends across intermediate periods (2000–2007, 2007–2015, and 2015–2023) are provided in Supplementary Tables S12–S15.
The gain–loss analysis of LULC transitions offers critical insight into the temporal dynamics of landscape change across the Babil and Al-Qadisiyah region. Over the full 2000–2023 period, CA exhibited a net gain of +514.85 km2, with substantial expansions observed between 2007–2015 (+204.87 km2) and 2015–2023 (+486.73 km2). This indicates continued agricultural intensification or land reclamation initiatives, following an initial contraction during 2000–2007 (–176.53 km2). HO also demonstrated a consistent positive trajectory, accumulating a total net gain of +128.65 km2. The most pronounced increase occurred in the first interval (+75.5 km2, 2000–2007), followed by continued, albeit smaller, gains thereafter. These trends suggest expansion of low-intensity vegetation cover or orchard-type agriculture. UL expanded steadily throughout all intervals, with a cumulative net gain of +70.43 km2, reflecting sustained urban growth. The largest increase occurred in 2015–2023 (+31.40 km2), likely driven by concentrated development in peri-urban areas. In contrast, WB experienced net instability, with a total net loss of –16.66 km2 over the study period. After an early expansion between 2000–2007 (+110.45 km2) possibly due to irrigation flooding or seasonal water retention WB subsequently contracted sharply during 2007–2015 (–20.79 km2) and 2015–2023 (–106.11 km2), highlighting vulnerability to hydrological stress or land conversion. BL showed the most substantial overall decline, with a cumulative net loss of –697.27 km2, primarily concentrated in 2015–2023 (–469.95 km2). These reductions likely reflect land rehabilitation, vegetative regrowth, or transformation into agricultural and built-up forms. Together, these patterns indicate a landscape undergoing substantial agricultural expansion and urban intensification, alongside ecological shifts in marginal and hydrologically sensitive areas Figure 4. The dynamics point to both socio-economic drivers and environmental responses shaping land cover transformation in central Iraq. A full breakdown of class-wise gain, loss, and net change across temporal intervals is provided in Supplementary Table S12. Disagreement analysis indicated moderate spatial uncertainty, with a total disagreement of 12.2%, driven primarily by allocation error (11.2%) rather than quantity mismatch (1.0%), underscoring the importance of spatial refinement in post-classification steps (Table S11).

3.2. Spatiotemporal Patterns of Vegetation

The annual trend analysis of FVC change revealed distinct spatial patterns of vegetation degradation and recovery across the study area between 2000 and 2023. Based on reclassified Sen’s slope values (% per decade), approximately 51.5% of the landscape exhibited strong vegetation recovery, with an average increase of +29.6% per decade. Additionally, 7.3% of the area showed moderate recovery, corresponding to an average trend of +7.3% per decade. These regions reflect successful regrowth, afforestation efforts, or favorable climatic conditions supporting vegetation expansion. Conversely, areas experiencing degradation were spatially limited yet ecologically important (Figure 5). Approximately 2.5% of the area experienced severe degradation, with a mean decline of –22.0% per decade, while 1.2% was moderately degraded, averaging –7.4% per decade (Table S16). These zones represent hotspots of vegetation stress due to persistent anthropogenic pressure, climatic drought, or land use transitions. Notably, 37.5% of the area was categorized as stable, indicating relatively unchanged vegetation conditions over the two-decade period (mean trend: +0.33%/decade).
The asymmetry between areas under strong recovery and those under degradation suggests a net greening trend in the region, though localized degradation persists. This aligns with similar findings in arid and semi-arid ecosystems where vegetation trends are heterogeneous and often driven by a combination of climatic and anthropogenic factors [68,69].

3.3. Spatiotemporal Patterns of Degradation

The composite degradation severity analysis, derived from the integrated assessment of spectral magnitude, degradation rate, and signal-to-noise ratio (DSNR), revealed that most affected areas were classified under Moderate Severity, covering 449.39 km2, followed by Low Severity with 207.65 km2 (Figure 6B). Only a small fraction (10.4 km2) exhibited High Severity, indicating that while degradation was widespread, extreme vegetation loss was relatively limited in spatial extent. Temporal stratification of the Year of Disturbance (YOD) layer into four decadal intervals (1990s, 2000s, 2010s, 2020s) highlighted a pronounced concentration of disturbance events during the 1990s, accounting for 534.24 km2, or most of all detected degradation (Figure 6A). In contrast, the 2000s and 2010s recorded minimal degradation activity (19.66 km2 and 16.94 km2, respectively), while a recovery was observed in the 2020s, with 96.18 km2 of degradation identified (Table S18). This pattern suggests that the most intense phase of long-term landscape transformation occurred in the early period of the time series, followed by a period of relative stability and then renewed degradation pressure in recent years. Together, these findings underscore both the intensity and timing of vegetation decline, with moderate degradation dominating and early 1990s disturbance waves shaping much of the current degraded landscape.

3.4. Degradation Attribution and Risk Prediction

3.4.1. The Influence of LUCC on Degradation

Overlaying severity classes with LULC transitions (2000–2023) showed that high-severity degradation was concentrated in transitions from Cultivated Agriculture to Bare Land (CA → BL) and from Cultivated Agriculture to Urban Land (CA → UL), indicating hotspots of anthropogenic disturbance and urban expansion. A substantial share of degradation also occurred within stable vegetated covers, notably CA → CA (309.53 km2 under Low–Moderate severity) and HO → HO (186.41 km2) (Table S19), consistent with productivity losses or soil quality decline without an overt land-cover change. Additional transitions such as CA → HO (25.21 km2) and CA → UL (2.83 km2) reflect shifts in agricultural practices and peri-urban growth. For interpretability, severity statistics were computed only for vegetated pixels (annual NDVI ≥ 0.10 or FVC ≥ 0.05); we did not summarize UL → UL or WB → WB combinations because NBR is not diagnostic over impervious or open-water surfaces and residual signals mostly arise from mixed 30-m edges. Overall, the patterns underscore degradation’s dual expression both through explicit land-cover transitions and in situ declines within persistent vegetated classes highlighting the need for management that targets visibly changing as well as chronically stressed areas.

3.4.2. The Influence of Climatic Drought Frequency and Joint Ecological Stress on Degradation

The analysis of climatic drought severity using the SPEI revealed consistent and spatially extensive drought stress across the study area (Figure 7). Pixel-wise percentage calculations showed that the mean duration under severe drought conditions (SPEI < –1.5) increased with the temporal scale of accumulation. Specifically, the 12-month scale (SPEI-12) exhibited the highest mean percentage of drought exposure at 7.44%, followed by 6.31% for SPEI-06 and 5.79% for SPEI-03 (Table S21). This trend reflects the cumulative nature of hydrological and soil moisture deficits over extended periods, which is particularly relevant in semi-arid environments where water retention and aquifer recharge are slow. Spatial variability was evident, with localized hotspots reaching up to 9.75% under SPEI-12-defined severe drought. In contrast, the minimum drought exposure remained above 3.97%, confirming that no area was completely exempt from long-term climatic stress. These findings are consistent with regional rainfall anomalies and evapotranspiration dynamics over the two-decade period.
To evaluate the ecological manifestation of climatic drought, joint stress layers were derived by overlaying severe drought conditions (SPEI < –1.5) with vegetation health stress (VHI < 40). The resulting joint indices showed considerably lower mean values, indicating that not all drought events translated into observable vegetation stress. The mean joint stress percentages were 1.30%, 1.09%, and 1.74% for SPEI03–VHI, SPEI06–VHI, and SPEI12–VHI, respectively. This discrepancy highlights the potential buffering capacity of vegetation or the role of anthropogenic interventions, such as irrigation, in mitigating short-term ecological impacts. To account for delayed vegetation responses, lagged joint stress layers were computed by offsetting VHI by one month. The lag-adjusted mean joint stress values slightly increased to 1.27% (SPEI03), 1.47% (SPEI06), and 1.49% (SPEI12). The increase was most notable at intermediate and longer accumulation timescales, reflecting delayed depletion of soil moisture and vegetation response inertia. These results underscore the importance of integrating climatic and biophysical indicators along with time-lag considerations to enhance drought impact assessments in dryland ecosystems. The modest yet consistent rise in stress following lag adjustment confirms the presence of ecological delay mechanisms, reinforcing the value of lag-integrated remote sensing indices in ecological drought diagnostics and land degradation monitoring frameworks.

3.4.3. Degradation Risk Probability Mapping

Validation, pixel-level probability outputs were mapped to produce a risk-probability surface (Figure 8). Probabilities were re-classified into Low (<0.40), Moderate (0.40–0.70) and High (>0.70) degradation-risk categories. Spatial aggregation showed that 38.5% of the landscape (≈6180.5 km2) falls under moderate risk, 51.7% (≈8280.4 km2) under low risk, and 9.7% (≈1553.7 km2) under high risk. This spatially explicit stratification supports targeted conservation planning by highlighting both areas already degraded and zones vulnerable to future decline under continuing climatic and anthropogenic pressures. On spatial block folds, drop-group permutation (conditional within LULC) indicated a larger AUC loss for the climate block than for the anthropogenic b l o c k   b a s e l i n e A U C = 0.638 ; Δ A U C climate = 0.065 ; Δ A U C anthro = 0.042 ;   d i f f e r e n c e = + 0.023 ,   95 % C I 0.042 ,   0.088 (Table S22). Grouped |SHAP| proportions were similar between blocks (climate = 0.472, anthropogenic = 0.528; difference = −0.056, 95% CI [−0.308, 0.196]) (Table S22). In leave-group-out ablation, removing anthropogenic predictors reduced AUC slightly more than removing climate b a s e l i n e   A U C = 0.614 ; Δ A U C drop ,   anthro = + 0.016   v s .   Δ A U C drop ,   climate = 0.006 ; d i f f e r e n c e = 0.022 ,   95 % C I 0.094 ,   0.049 (Table S23). Taken together, these diagnostics support comparable aggregate contributions at regional scale, with context dependence across irrigated versus non-irrigated landscapes.
Summarizing risk classes by hydrological proximity shows that the share of Moderate + High risk declines with distance from channels (≤2 km: 36.7%; 2–10 km: 31.9%; >10 km: 29.3%; Table S24). Low-risk area increases from 63.3% (≤2 km) to 70.7% (>10 km), while High risk is concentrated both near channels (9.8%) and in distal dryland belts (12.2%), reflecting irrigated corridors versus rain-fed/desert margins.
SHAP-based interpretation revealed the internal logic of the XGBoost classifier, as illustrated in Figure 9. Panel A presents the ranked distribution of main effect contributions. Among all predictors, SPEI-06 (mid-term drought) and SPEI-03 (short-term drought) emerged as the most influential drivers of degradation probability. Higher drought severity values (orange hue) were consistently associated with increased model-estimated disturbance likelihood. Other significant contributors included agricultural intensity, distance to water bodies (D-WB), and SPEI-12, reflecting the compounded effects of land-use pressure and long-term climatic stress. Predictors with minimal influence such as categorical land cover types (CA, HO, UL)—showed near-zero contribution due to mutual exclusivity and limited variation across the sample. Panel B highlights the top ten pairwise SHAP interaction strengths, revealing that degradation patterns are shaped not just by individual drivers but by their combined effects. The strongest synergy was observed between distance to rivers and SPEI-06 (interaction value: 0.038), underscoring the combined influence of hydrological proximity and mid-term drought stress. Equally strong interactions were found between distance to rivers and roads (0.032), and SPEI-06 with Joint-12, an integrated ecological stress index (0.032). Additional co-dependencies included SPEI-03 with SPEI-06 (0.031) and agricultural intensity with distance to rivers (0.029), reflecting how degradation likelihood is shaped by the synergy between anthropogenic pressure and climatic variability. These findings confirm that the model encodes ecologically grounded and interpretable relationships, offering transparent insight into the multifactorial nature of vegetation degradation in drought-prone landscapes.

3.4.4. Model Performance

Model performance was evaluated using an independent test set comprising 30% of samples, geographically buffered by 5 km from training data to reduce spatial autocorrelation (Section 2). On this held-out data, the XGBoost classifier achieved 79.2% overall accuracy and a Cohen’s κ of 0.58, indicating moderate-to-substantial agreement. Sensitivity for degraded areas was 0.75, specificity for stable vegetation 0.83, precision for degraded class 0.82, and F1-score 0.78, showing reliable degradation detection. ROC analysis yielded a test AUC of 0.884, confirming strong discriminative power (Table 4). Training set metrics were slightly higher, with 83.5% accuracy, κ of 0.67, sensitivity 0.80, specificity 0.87, and AUC 0.925. AUC difference (~0.04) indicated acceptable generalization, with overfitting mitigated by hyperparameter tuning and spatial five-fold cross-validation (mean CV AUC 0.911 ± 0.006). McNemar’s χ2 test was significant (p < 0.001), confirming non-random misclassification. Supplementary Figure S2 displays ROC curves for both training and test sets, annotated with optimal thresholds (Youden index) and multiple decision cutoffs to aid interpretability across application scenarios. Probability outputs were mapped to generate a degradation risk surface (Figure 8), classified as Low (<0.40), Moderate (0.40–0.70), and High (>0.70) risk. Spatially, 38.5% (~6180.5 km2) of the landscape was moderate risk, 51.7% (~8280.4 km2) low risk, and 9.7% (~1553.7 km2) high risk, supporting targeted conservation planning.
The XGBoost model demonstrated strong predictive performance, with an overall accuracy of ~79%, comparable to other studies in arid regions. For instance, a similar XGBoost model in Turkmenistan reached 96% accuracy [70]. While this model’s performance was slightly lower, it reflects the more complex task of pre-dicting degradation across heterogeneous landscapes. The model balanced sensitivity and specificity, which is crucial for accurately distinguishing degraded from stable areas and aligns with literature showing that tree-based models outperform simpler algo-rithms in capturing complex land degradation processes [71,72]. The XGBoost model’s ability to handle imbalanced data and integrate diverse predictors was also a key strength, as seen in other environmental applications [21].

4. Discussion

4.1. Land Degradation Patterns in Iraq

The observed land-use changes in Babil and Al-Qadisiyah are consistent with broader Middle Eastern dryland dynamics. Urban expansion is evident and accords with regional reports of rapidly growing built-up areas displacing agricultural or natural land [14,73]. CA shows a brief early-period contraction (2000–2007) followed by net expansion by 2023, while bare land declines overall and HO register modest gains (Tables S1 and S12). Local conversions of CA → BL and CA → UL do occur, and mark localized degradation or peri-urban encroachment, but they do not represent a sustained basin-wide loss of cropland. These patterns fit regional evidence that degradation risk concentrates where vegetation is sparse and water management is weak [18], whereas pockets of resilience and recovery including slight increases in HO are associated with targeted management and reclamation efforts reported in neighboring countries [74,75].
Vegetation index analysis highlights the interplay between climatic droughts and land use changes. Prolonged droughts led to significant drops in vegetation health (NDVI/VHI) and were often followed by land cover changes. While short-term droughts were sometimes mitigated by irrigation or natural resilience, prolonged droughts (indicated by severely low SPEI values) resulted in more severe vegetation decline and land use shifts. This is consistent with findings in other drylands, where extended moisture deficits cause lasting ecological damage [9,10]. The study revealed that although drought affected the entire area over the past two decades, extreme drought-vegetation stress overlaps were limited, suggesting some buffering capacity or human interventions. The introduction of a one-month lag showed that vegetation response to drought is often delayed, supporting findings from North China where vegetation indices reflected cumulative rainfall deficits with time delays [22]. This highlights that, in Iraqi agro-ecosystems, droughts act as catalysts for land degradation, especially when compounded by poor land management practices. Conversely, lands with adaptive strategies (e.g., crop switching, irrigation) were more resilient. Areas with adaptive responses showed better resilience, as indicated by transitions from Cropland to HO (drought-tolerant horticulture). This confirms the findings from Pakistan’s rangelands, where human activities were more significant than climate in driving land degradation [76].

4.2. Dominant Predictors of Degradation and the Risk in the Future

The model identified mid-term drought (SPEI-06) and short-term drought (SPEI-03) as dominant predictors of degradation, confirming that water balance anomalies are critical drivers of vegetation loss in drylands [77,78]. Distance to water bodies and agricultural intensity were also significant, indicating that areas farther from irrigation sources and under intensive land use were more prone to degradation. SHAP analysis provided valuable insights into the model’s decision-making process. It confirmed that increasing drought severity raised the predicted probability of degradation, while proximity to water bodies reduced this likelihood. Interaction effects, particularly between prolonged drought (SPEI-06) and poor hydrological access, had a synergistic impact on degradation risk, supporting the intuitive understanding that droughts have more severe consequences in areas distant from water sources [79]. These results align with studies in similar environments, where proximity to water plays a critical role in drought resilience [80,81]. Additionally, the interaction between infrastructure (distance to roads) and drought further underscores the vulnerability of areas on the fringes of human support networks [82].
This study has some limitations. First, remote sensing indices like NDVI and VHI can be affected by factors unrelated to degradation, such as soil properties or atmospheric conditions. While LandTrendr helped focus on persistent vegetation losses, subtle degradation, such as soil quality decline, may remain undetected. Classification errors in land cover mapping, particularly between spectrally similar classes, could also influence transition analysis. The SPEI index, based on gridded climate data, may not capture local micro-climatic variations. Future research should integrate ground-level data on irrigation and groundwater usage to improve drought modeling. Although the XGBoost model performed well, its moderate Cohen’s κ suggests some degradation patterns remain difficult to predict, possibly due to unmodeled factors. Incorporating spatially explicit features or spatio-temporal models could address this. Higher resolution data (e.g., Sentinel-2) and newer sensors (e.g., LiDAR) could capture finer-scale degradation signals. Future studies should also extend the analysis using climate projections to assess future degradation risks under climate change. Continued field validation is necessary, particularly in areas of low prediction confidence, to ensure model accuracy. Our methodology, combining remote sensing, drought analysis, LULC change detection, and machine learning, offers a comprehensive framework for land degradation assessment. Future work can enhance this approach by incorporating socio-economic data to refine degradation of drivers and support land restoration and climate adaptation strategies.

4.3. Model Performance—Strengths, Limitations, and Implications

Our classifier’s performance on a spatially independent, buffered hold-out indicates that the model captures meaningful signal rather than local spatial autocorrelation. This is consistent with the literature showing that tree-based ensembles remain competitive for heterogeneous drylands where relationships are non-linear and interactions are common [83,84]. The combination of spatial block CV, buffered test splits, and bootstrap ensembling provides complementary checks on generalization and epistemic uncertainty. Model interpretability Via grouped SHAP and conditional drop-group permutation further supports ecological plausibility drought metrics carry substantial influence, modulated by hydrological proximity and land-use pressure rather than relying on opaque feature rankings. There are, however, clear limitations. First, climate predictors were up-scaled to 30 m only for co-registration; their true support remains coarse, which can dampen fine-scale contrasts and propagate scale mismatch. Second, labels inherit uncertainty from the disturbance workflow (LandTrendr, masks, MMU), which may blur the decision boundary in irrigation mosaics or transient fallows. Third, several anthropogenic covariates (e.g., OSM distances, population) are proxies; unobserved management (groundwater pumping, canal operations, grazing) likely explains residual error. Finally, while the model transported well within our domain, transferability to other basins should be tested before operational use. Practically, these strengths and caveats suggest two uses: (i) relative risk zoning the map is most reliable for prioritizing surveillance and field verification; and (ii) hypothesis generation SHAP patterns help target drought–hydrology interactions for monitoring. Future work could incorporate sub-pixel metrics (e.g., SMA/NDFI), seasonally smoothed inputs (harmonic/moving-window composites), and physiographic stratification (watersheds/eco-hydro zones) to further stabilize predictions and enhance spatial coherence [85].

4.4. Implications for Monitoring and Management

The integrated outputs support tiered monitoring: trend and disturbance layers flag legacy impacts; transition matrices prioritise conversion fronts (e.g., CA → BL, CA → UL); the risk map and hydro-stratified summaries identify zones where irrigation access and restoration would yield the largest marginal risk reduction. Agencies can combine the high-risk (>0.70) surface with proximity to infrastructure to stage cost-effective interventions (canal maintenance, salinity control, shelterbelts) [86]. At basin scale, the climate-led signal argues for coupling drought planning (multi-timescale SPEI triggers) with demand management (cropping calendars, deficit irrigation) [87]. Two extensions are immediate. First, incorporate sub-pixel metrics (SMA/NDFI) to refine Sparse Vegetation Vs. Bare Soil discrimination in dune–alluvial mosaics. Second, add seasonal smoothing (harmonic or moving-window metrics) to stabilise trend and disturbance inputs in heavily cropped corridors. Both would test sensitivity of attribution while preserving the current block-wise framework.

5. Conclusions

This study provides a spatially explicit account of vegetation degradation risk in Babil–Al-Qadisiyah (2000–2023) by linking long-term trends, disturbance segmentation, land-use transitions, and model-based attribution. The central hypothesis—multi-timescale climatic water deficits exert a stronger effect on degradation risk than anthropogenic pressures, conditional on hydrological connectivity—was evaluated using complementary diagnostics (conditional permutation ΔAUC, grouped SHAP, leave-group-out ablation) under spatial block cross-validation. Three findings stand out. (i) Driver attribution indicates a climate-led signal: conditional permutation produces a larger mean loss in AUC when the climate block is perturbed, whereas grouped SHAP shares are near-balanced with a slight anthropogenic share, and ablation differences are small within uncertainty. (ii) The XGBoost model attains strong discrimination (test AUC = 0.884) and maps 9.7% of the landscape as high risk (>0.70), with risk concentrated away from perennial water bodies, underscoring the buffering role of irrigation connectivity. (iii) LULC dynamics show net expansion of cultivated agriculture (+515 km2), herbaceous/orchards (+129 km2), and urban land (+70 km2) alongside a marked contraction of bare land (−697 km2); trend mapping from MK–Sen indicates recovery across 51.5% of the area and severe decline on 2.5%. Methodologically, this work integrates peak-season LandTrendr on NBR for disturbance timing/severity, matrix-corrected transition accounting, and block-wise attribution of climate versus anthropogenic influence, an explicit comparison rarely reported in canal-fed drylands. Uncertainty is quantified through a bootstrap ensemble, delivered alongside the risk surface. Several scope conditions apply. TerraClimate fields (4 km) were resampled only for co-registration; climate effects should be interpreted at their native spatial support. Despite peak-season compositing, residual phenological aliasing may persist, and change estimates inherit classification uncertainty (reported with matrix correction and disagreement metrics). Future extensions include sub-pixel metrics (SMA/NDFI) to refine sparse vegetation versus bare soil discrimination, seasonal smoothing of inputs, and physiographic stratification for performance reporting; incorporating irrigation/groundwater observations would further strengthen inference. Practically, the outputs support tiered monitoring and targeting transition matrices highlight conversion fronts (CA → BL, CA → UL), while the risk map stratified by hydrological proximity prioritizes zones where restoration and water management measures (canal maintenance, salinity control, adaptive cropping) can most effectively reduce degradation risk. The framework is transferable to other irrigated drylands where disentangling climate stress from land-use pressure is central to desertification management.

Supplementary Materials

The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/rs17193343/s1. References [11,39,45,57,58,88,89,90,91,92,93,94,95,96,97,98,99] are cited in the supplementary materials.

Author Contributions

Conceptualization: N.A.-T.; Data curation, N.A.-T.; Formal analysis, N.A.-T. and F.S.; Funding acquisition, J.Z.; Investigation, N.A.-T.; Methodology, N.A.-T.; Software, N.A.-T.; Validation, N.A.-T.; Visualization, N.A.-T. and F.S.; Writing—original draft, N.A.-T.; Writing—review & editing, Z.X., F.S., K.M., X.L. and J.Z. All authors have read and agreed to the published version of the manuscript.

Funding

This research is funded by the National Natural Science Foundation of China (grant number: 42361144885) and Beijing Forestry University (grant number: BLRC2023B09). And the APC was funded by the National Natural Science Foundation of China and the 5.5 Engineering Research & Innovation Team Project of Beijing Forestry University.

Institutional Review Board Statement

This research did not involve human or animal subjects; therefore, formal ethical approval was not required. The study strictly adheres to general ethical principles, and the authors are committed to upholding the highest standards of ethical research conduct. Any potential conflicts of interest that could have influenced the ethical conduct of this research have been declared.

Informed Consent Statement

Informed consent was obtained from all participants involved in this study. Participants were provided with detailed information about the research objectives, procedures, potential risks, and benefits before agreeing to participate. They were assured that their participation was voluntary, and they had the right to withdraw from the study at any time without facing any consequences. All participants were informed about the confidentiality measures in place to protect their identity and personal information.

Data Availability Statement

Data collected during the study will be used solely for research purposes and will be securely stored. This study was conducted in accordance with ethical standards and guidelines, and participants were encouraged to ask questions and seek clarification at any stage of the research process. If you have any further questions or concerns regarding the consent process, please contact nawar1992m@gmail.com.

Acknowledgments

N.A.T., Z.X., F.S., K.M., X.L. and J.X. acknowledge the support of the School of Soil and Water Conservation, Beijing Forestry University, for help and platforms for this research.

Conflicts of Interest

The authors declare no conflicts of interest that could have influenced the ethical conduct of this research.

References

  1. Singh, A.K.; Shah, S.K.; Pandey, U.; Deeksha; Thomte, L.; Rahman, T.W.; Mehrotra, N.; Singh, D.S.; Kotlia, B.S. Vegetation Index (NDVI) Reconstruction from Western Himalaya through Dendrochronological Analysis of Cedrus Deodara. Theor. Appl. Clim. 2024, 155, 1713–1727. [Google Scholar] [CrossRef]
  2. Burger, R.; Aouizerats, B.; den Besten, N.; Guillevic, P.; Catarino, F.; van der Horst, T.; Jackson, D.; Koopmans, R.; Ridderikhoff, M.; Robson, G.; et al. The Biomass Proxy: Unlocking Global Agricultural Monitoring through Fusion of Sentinel-1 and Sentinel-2. Remote Sens. 2024, 16, 835. [Google Scholar] [CrossRef]
  3. Anees, S.A.; Mehmood, K.; Khan, W.R.; Sajjad, M.; Alahmadi, T.A.; Alharbi, S.A.; Luo, M. Integration of Machine Learning and Remote Sensing for above Ground Biomass Estimation through Landsat-9 and Field Data in Temperate Forests of the Himalayan Region. Ecol. Inform. 2024, 82, 102732. [Google Scholar] [CrossRef]
  4. Khan, W.R.; Nazre, M.; Akram, S.; Anees, S.A.; Mehmood, K.; Ibrahim, F.H.; Al Edrus, S.S.O.; Latiff, A.; Fitri, Z.A.; Yaseen, M.; et al. Assessing the Productivity of the Matang Mangrove Forest Reserve: Review of One of the Best-Managed Mangrove Forests. Forests 2024, 15, 747. [Google Scholar] [CrossRef]
  5. Khan, W.R.; Giani, M.; Bevilacqua, S.; Anees, S.A.; Mehmood, K.; Nazre, M.; Haddy, A.A.B.A.; Median, A.N.B.A.; Bin Bujang, J.; Mohamad-Ismail, F.-N.; et al. Derivation of Allometric Equations and Carbon Content Estimation in Mangrove Forests of Malaysia. Environ. Sustain. Indic. 2025, 26, 100618. [Google Scholar] [CrossRef]
  6. Mehmood, K.; Anees, S.A.; Rehman, A.; Rehman, N.U.; Muhammad, S.; Shahzad, F.; Liu, Q.; Alharbi, S.A.; Alfarraj, S.; Ansari, M.J. Assessment of Climatic Influences on Net Primary Productivity along Elevation Gradients in Temperate Ecoregions. Trees For. People 2024, 18, 100657. [Google Scholar] [CrossRef]
  7. Khan, J.; Wang, P.; Xie, Y.; Wang, L.; Li, L. Mapping MODIS LST NDVI Imagery for Drought Monitoring in Punjab Pakistan. IEEE Access 2018, 6, 19898–19911. [Google Scholar] [CrossRef]
  8. Stagge, J.H.; Tallaksen, L.M.; Xu, C.Y.; Van Lanen, H.A.J. Standardized Precipitation-Evapotranspiration Index (SPEI): Sensitivity to Potential Evapotranspiration Model and Parameters. In Proceedings of the 7th World Flow Regimes from International and Experimental Network Data-Water Conference, FRIEND-Water 2014, Montpellier, France, 7–10 October 2014; Volume 363. [Google Scholar]
  9. Chaparro, D.; Jagdhuber, T.; Piles, M.; Jonard, F.; Fluhrer, A.; Vall-llossera, M.; Camps, A.; López-Martínez, C.; Fernández-Morán, R.; Baur, M.; et al. Vegetation Moisture Estimation in the Western United States Using Radiometer-Radar-Lidar Synergy. Remote Sens. Environ. 2024, 303, 113993. [Google Scholar] [CrossRef]
  10. Wang, M.; Ciais, P.; Fensholt, R.; Brandt, M.; Tao, S.; Li, W.; Fan, L.; Frappart, F.; Sun, R.; Li, X.; et al. Satellite Observed Aboveground Carbon Dynamics in Africa during 2003–2021. Remote Sens. Environ. 2024, 301, 113927. [Google Scholar] [CrossRef]
  11. Bento, V.A.; Trigo, I.F.; Gouveia, C.M.; DaCamara, C.C. Contribution of Land Surface Temperature (TCI) to Vegetation Health Index: A Comparative Study Using Clear Sky and All-Weather Climate Data Records. Remote Sens. 2018, 10, 1324. [Google Scholar] [CrossRef]
  12. Chere, Z.; Debalke, D.B. Modeling Agricultural Drought Based on the Earth Observation-Derived Standardized Precipitation Evapotranspiration Index and Vegetation Health Index in the Northeastern Highlands of Ethiopia. Nat. Hazards 2024, 120, 3127–3151. [Google Scholar] [CrossRef]
  13. Hang, Q.; Guo, H.; Meng, X.; Wang, W.; Cao, Y.; Liu, R.; De Maeyer, P.; Wang, Y. Optimizing the Vegetation Health Index for Agricultural Drought Monitoring: Evaluation and Application in the Yellow River Basin. Remote Sens. 2024, 16, 4507. [Google Scholar] [CrossRef]
  14. Selmy, S.A.H.; Kucher, D.E.; Mozgeris, G.; Moursy, A.R.A.; Jimenez-Ballesta, R.; Kucher, O.D.; Fadl, M.E.; Mustafa, A. rahman A. Detecting, Analyzing, and Predicting Land Use/Land Cover (LULC) Changes in Arid Regions Using Landsat Images, CA-Markov Hybrid Model, and GIS Techniques. Remote Sens. 2023, 15, 5522. [Google Scholar] [CrossRef]
  15. Salih, N.M.; Nori, A.F.; Alatta, H.J.; Khalaf, S.Z. Comparative Analysis for Dust and Sand Storms In Iraq: A Survey. J. Port. Sci. Res. 2024, 6, 28–31. [Google Scholar] [CrossRef]
  16. Hussain, K.; Badshah, T.; Mehmood, K.; Rahman, A.U.; Shahzad, F.; Anees, S.A.; Khan, W.R.; Yujun, S. Comparative Analysis of Sensors and Classification Algorithms for Land Cover Classification in Islamabad, Pakistan. Earth Sci. Inf. 2025, 18, 212. [Google Scholar] [CrossRef]
  17. Mehmood, K.; Anees, S.A.; Luo, M.; Akram, M.; Zubair, M.; Khan, K.A.; Khan, W.R. Assessing Chilgoza Pine (Pinus Gerardiana) Forest Fire Severity: Remote Sensing Analysis, Correlations, and Predictive Modeling for Enhanced Management Strategies. Trees For. People 2024, 16, 100521. [Google Scholar] [CrossRef]
  18. Al-Obaidi, J.R.; Allawi, M.Y.; Al-Taie, B.S.; Alobaidi, K.H.; Al-Khayri, J.M.; Abdullah, S.; Ahmad-Kamil, E.I. The Environmental, Economic, and Social Development Impact of Desertification in Iraq: A Review on Desertification Control Measures and Mitigation Strategies. Environ. Monit. Assess. 2022, 194, 1–18. [Google Scholar] [CrossRef]
  19. Liu, Y.; Qiu, H.; Wang, N.; Yang, D.; Zhao, K.; Yang, G.; Huangfu, W.; Luo, W. Thermokarst Disturbance Responses to Climate Change across the Circumpolar Permafrost Regions from 1990 to 2023. Geosci. Front. 2025, 16, 102147. [Google Scholar] [CrossRef]
  20. Yan, F.; He, B.; Lyne, V.; Fan, R.; Cui, Y.; Wang, X.; Fu, D.; Meadows, M.; Wilson, J.; Chen, Z. Global Coastal Water Clarity Has Increased Due to Human Intervention. Commun. Earth Environ. 2025, 6, 641. [Google Scholar] [CrossRef]
  21. Hao, Q.; Zhang, T.; Cheng, X.; He, P.; Zhu, X.; Chen, Y. GIS-Based Non-Grain Cultivated Land Susceptibility Prediction Using Data Mining Methods. Sci. Rep. 2024, 14, 4433. [Google Scholar] [CrossRef]
  22. Zhang, Z.; Wang, Z.; Lai, H.; Wang, F.; Li, Y.; Feng, K.; Qi, Q.; Di, D. Lag Time and Cumulative Effects of Climate Factors on Drought in North China Plain. Water 2023, 15, 3428. [Google Scholar] [CrossRef]
  23. Ali, S.; Haixing, Z.; Qi, M.; Liang, S.; Ning, J.; Jia, Q.; Hou, F. Monitoring Drought Events and Vegetation Dynamics in Relation to Climate Change over Mainland China from 1983 to 2016. Environ. Sci. Pollut. Res. 2021, 28, 21910–21925. [Google Scholar] [CrossRef]
  24. Mehmood, K.; Anees, S.A.; Muhammad, S.; Hussain, K.; Shahzad, F.; Liu, Q.; Ansari, M.J.; Alharbi, S.A.; Khan, W.R. Analyzing Vegetation Health Dynamics across Seasons and Regions through NDVI and Climatic Variables. Sci. Rep. 2024, 14, 11775. [Google Scholar] [CrossRef]
  25. Jiang, L.; Liu, B.; Yuan, Y. Quantifying Vegetation Vulnerability to Climate Variability in China. Remote Sens. 2022, 14, 3491. [Google Scholar] [CrossRef]
  26. Shahzad, F.; Mehmood, K.; Hussain, K.; Haidar, I.; Anees, S.A.; Muhammad, S.; Ali, J.; Adnan, M.; Wang, Z.; Feng, Z. Comparing Machine Learning Algorithms to Predict Vegetation Fire Detections in Pakistan. Fire Ecol. 2024, 20, 57. [Google Scholar] [CrossRef]
  27. Osman, M. Kurdistan Region of Iraq Population Analysis Report; Kurdistan Region Statistics Office in Kurdistan Regional Government: Erbil, Irak, 2021.
  28. Al-Ansari, N. Topography and Climate of Iraq. J. Earth Sci. Geotech. Eng. 2020, 11, 1–13. [Google Scholar] [CrossRef]
  29. Al-Khalidi, J.; Dima, M.; Stefan, S. Large-Scale Modes Impact on Iraq Climate Variability. Theor. Appl. Clim. 2018, 133, 179–190. [Google Scholar] [CrossRef]
  30. Ali, S.H.; Qubaa, A.R.; Al-Khayat, A.B.M. Climate Change and Its Potential Impacts on Iraqi Environment: Overview. In Proceedings of the IOP Conference Series: Earth and Environmental Science, Makassar, Indonesia, 27 July 2024; Volume 1300. [Google Scholar]
  31. Su, Y.; Cui, Y.J.; Dupla, J.C.; Canou, J. Soil-Water Retention Behaviour of Fine/Coarse Soil Mixture with Varying Coarse Grain Contents and Fine Soil Dry Densities. Can. Geotech. J. 2022, 59, 291–299. [Google Scholar] [CrossRef]
  32. Wang, N.; Wu, Q.; Gui, Y.; Hu, Q.; Li, W. Cross-Modal Segmentation Network for Winter Wheat Mapping in Complex Terrain Using Remote-Sensing Multi-Temporal Images and DEM Data. Remote Sens. 2024, 16, 1775. [Google Scholar] [CrossRef]
  33. Ali, J.; Haoran, W.; Mehmood, K.; Hussain, W.; Iftikhar, F.; Shahzad, F.; Hussain, K.; Qun, Y.; Zhongkui, J. Remote Sensing and Integration of Machine Learning Algorithms for Above-Ground Biomass Estimation in Larix Principis-Rupprechtii Mayr Plantations: A Case Study Using Sentinel-2 and Landsat-9 Data in Northern China. Front. Environ. Sci. 2025, 13, 1577298. [Google Scholar] [CrossRef]
  34. Bar, S.; Parida, B.R.; Pandey, A.C. Landsat-8 and Sentinel-2 Based Forest Fire Burn Area Mapping Using Machine Learning Algorithms on GEE Cloud Platform over Uttarakhand, Western Himalaya. Remote Sens. Appl. 2020, 18, 100324. [Google Scholar] [CrossRef]
  35. Abatzoglou, J.T.; Dobrowski, S.Z.; Parks, S.A.; Hegewisch, K.C. TerraClimate, a High-Resolution Global Dataset of Monthly Climate and Climatic Water Balance from 1958-2015. Sci. Data 2018, 5, 170191. [Google Scholar] [CrossRef]
  36. Gondo, R.; Kolawole, O.D. Dynamics of Land Use and Land Cover Change in the Distal Okavango Delta, Botswana. Geol. Ecol. Landsc. 2024, 1–18. [Google Scholar] [CrossRef]
  37. Gu, D.; Andreev, K.; Dupre, M.E. Major Trends in Population Growth Around the World. China CDC Wkly. 2021, 3, 604–613. [Google Scholar] [CrossRef]
  38. Pesaresi, M.; Schiavina, M.; Politis, P.; Freire, S.; Krasnodębska, K.; Uhl, J.H.; Carioli, A.; Corbane, C.; Dijkstra, L.; Florio, P. Advances on the Global Human Settlement Layer by Joint Assessment of Earth Observation and Population Survey Data. Int. J. Digit. Earth 2024, 17, 2390454. [Google Scholar] [CrossRef]
  39. Mehmood, K.; Anees, S.A.; Muhammad, S.; Albasher, G.; Shahzad, F.; Hussain, K.; Liu, Q.; Ayub, R.; Khan, W.R. Spatial and Temporal Vegetation Dynamics from 2000 to 2023 in the Western Himalayan Regions. Stoch. Environ. Res. Risk Assess. 2025, 39, 2309–2330. [Google Scholar] [CrossRef]
  40. Mehmood, K.; Anees, S.A.; Muhammad, S.; Shahzad, F.; Liu, Q.; Khan, W.R.; Shrahili, M.; Ansari, M.J.; Dube, T. Machine Learning and Spatio Temporal Analysis for Assessing Ecological Impacts of the Billion Tree Afforestation Project. Ecol. Evol. 2025, 15, e70736. [Google Scholar] [CrossRef]
  41. Huang, E.; Zhu, G.; Meng, G.; Wang, Y.; Chen, L.; Miao, Y.; Wang, Q.; Shi, X.; Zhao, L.; Wang, Q. Historical Dataset of Reservoir Construction in Arid Regions. Sci. Data 2025, 12, 1428. [Google Scholar] [CrossRef]
  42. Moulds, S.; Buytaert, W.; Mijic, A. An Open and Extensible Framework for Spatially Explicit Land Use Change Modelling: The Lulcc R Package. Geosci. Model. Dev. 2015, 8, 3215–3229. [Google Scholar] [CrossRef]
  43. Baghel, S.; Kothari, M.K.; Tripathi, M.P.; Singh, P.K.; Bhakar, S.R.; Dave, V.; Jain, S.K. Spatiotemporal LULC Change Detection and Future Prediction for the Mand Catchment Using MOLUSCE Tool. Env. Earth Sci. 2024, 83, 66. [Google Scholar] [CrossRef]
  44. Li, R.; Qi, X.; Chen, L.; Zhu, G.; Meng, G.; Wang, Y.; Huang, E.; Jiao, Y.; Wang, Q.; Li, W. Hydrological Processes in Continental Valley Basins: Evidence from Water Stable Isotopes. Catena 2025, 259, 109314. [Google Scholar] [CrossRef]
  45. Pontius, R.G.; Millones, M. Death to Kappa: Birth of Quantity Disagreement and Allocation Disagreement for Accuracy Assessment. Int. J. Remote Sens. 2011, 32, 4407–4429. [Google Scholar] [CrossRef]
  46. García-Álvarez, D.; Camacho Olmedo, M.T.; Van Delden, H.; Mas, J.F.; Paegelow, M. Comparing the Structural Uncertainty and Uncertainty Management in Four Common Land Use Cover Change (LUCC) Model Software Packages. Environ. Model. Softw. 2022, 153, 105411. [Google Scholar] [CrossRef]
  47. Shahzad, F.; Mehmood, K.; Anees, S.A.; Adnan, M.; Muhammad, S.; Haidar, I.; Ali, J.; Hussain, K.; Feng, Z.; Khan, W.R. Advancing Forest Fire Prediction: A Multi-Layer Stacking Ensemble Model Approach. Earth Sci. Inf. 2025, 18, 270. [Google Scholar] [CrossRef]
  48. Anees, S.A.; Mehmood, K.; Rehman, A.; Rehman, N.U.; Muhammad, S.; Shahzad, F.; Hussain, K.; Luo, M.; Alarfaj, A.A.; Alharbi, S.A.; et al. Unveiling Fractional Vegetation Cover Dynamics: A Spatiotemporal Analysis Using MODIS NDVI and Machine Learning. Environ. Sustain. Indic. 2024, 24, 100485. [Google Scholar] [CrossRef]
  49. Hijmans, R.J. Terra: Spatial Data Analysis. R Package Version 1.8-67. 2025. Available online: https://CRAN.R-project.org/package=terra (accessed on 26 June 2025).
  50. Zhang, W.; Wei, F.; Horion, S.; Fensholt, R.; Forkel, M.; Brandt, M. Global Quantification of the Bidirectional Dependency between Soil Moisture and Vegetation Productivity. Agric. Meteorol. 2022, 313, 108735. [Google Scholar] [CrossRef]
  51. Kennedy, R.E.; Yang, Z.; Gorelick, N.; Braaten, J.; Cavalcante, L.; Cohen, W.B.; Healey, S. Implementation of the LandTrendr Algorithm on Google Earth Engine. Remote Sens. 2018, 10, 691. [Google Scholar] [CrossRef]
  52. Gelabert, P.J.; Rodrigues, M.; de la Riva, J.; Ameztegui, A.; Sebastià, M.T.; Vega-Garcia, C. LandTrendr Smoothed Spectral Profiles Enhance Woody Encroachment Monitoring. Remote Sens. Environ. 2021, 262, 112521. [Google Scholar] [CrossRef]
  53. Sun, H.; Ma, X.; Liu, Y.; Zhou, G.; Ding, J.; Lu, L.; Wang, T.; Yang, Q.; Shu, Q.; Zhang, F. A New Multiangle Method for Estimating Fractional Biocrust Coverage from Sentinel-2 Data in Arid Areas. IEEE Trans. Geosci. Remote Sens. 2024, 62, 1–15. [Google Scholar] [CrossRef]
  54. Foga, S.; Scaramuzza, P.L.; Guo, S.; Zhu, Z.; Dilley, R.D., Jr.; Beckmann, T.; Schmidt, G.L.; Dwyer, J.L.; Hughes, M.J.; Laue, B. Cloud detection algorithm comparison and validation for operational Landsat data products. Remote Sens. Environ. 2017, 194, 379–390. [Google Scholar] [CrossRef]
  55. Pasquarella, V.J.; Arévalo, P.; Bratley, K.H.; Bullock, E.L.; Gorelick, N.; Yang, Z.; Kennedy, R.E. Demystifying LandTrendr and CCDC Temporal Segmentation. Int. J. Appl. Earth Obs. Geoinf. 2022, 110, 102806. [Google Scholar] [CrossRef]
  56. Qiu, D.; Liang, Y.; Shang, R.; Chen, J.M. Improving LandTrendr Forest Disturbance Mapping in China Using Multi-Season Observations and Multispectral Indices. Remote Sens. 2023, 15, 2381. [Google Scholar] [CrossRef]
  57. Stagge, J.H.; Tallaksen, L.M.; Gudmundsson, L.; Van Loon, A.F.; Stahl, K. Response to Comment on “Candidate Distributions for Climatological Drought Indices (SPI and SPEI)”. Int. J. Climatol. 2016, 36, 2132–2138. [Google Scholar] [CrossRef]
  58. Vicente-Serrano, S.M.; Beguería, S. Comment on “Candidate Distributions for Climatological Drought Indices (SPI and SPEI)” by James, H. Stagge et al. Int. J. Climatol. 2016, 36, 2120–2131. [Google Scholar] [CrossRef]
  59. Liu, W.; Wang, J.; Zuo, H.; Fu, Z.; Xiao, W.; Cui, Y.; Zhou, Z. Spatiotemporal Distribution and Variation Characteristics of Convective Activities in Different Climate Zones in Northern China Based on 25 Years of Satellite Observations. Int. J. Climatol. 2025, 45, e8908. [Google Scholar] [CrossRef]
  60. Gong, C.; Yue, X.; Liao, H.; Ma, Y. A Humidity-Based Exposure Index Representing Ozone Damage Effects on Vegetation. Environ. Res. Lett. 2021, 16, 044030. [Google Scholar] [CrossRef]
  61. Xiong, Y.; Dai, Z.; Long, C.; Liang, X.; Lou, Y.; Mei, X.; Nguyen, B.A.; Cheng, J. Machine Learning-Based Examination of Recent Mangrove Forest Changes in the Western Irrawaddy River Delta, Southeast Asia. Catena 2024, 234, 107601. [Google Scholar] [CrossRef]
  62. Feng, K.; Wang, T.; Liu, S.; Kang, W.; Chen, X.; Guo, Z.; Zhi, Y. Monitoring Desertification Using Machine-Learning Techniques with Multiple Indicators Derived from MODIS Images in Mu Us Sandy Land, China. Remote Sens. 2022, 14, 2663. [Google Scholar] [CrossRef]
  63. Lundberg, S.M.; Erion, G.G.; Lee, S.-I. Consistent individualized feature attribution for tree ensembles. arXiv 2019, arXiv:1802.03888. [Google Scholar] [CrossRef]
  64. DiCiccio, T.J.; Efron, B. Bootstrap Confidence Intervals. Stat. Sci. 1996, 11, 189–228. [Google Scholar] [CrossRef]
  65. Tibbe, T.D.; Montoya, A.K. Correcting the Bias Correction for the Bootstrap Confidence Interval in Mediation Analysis. Front. Psychol. 2022, 13, 810258. [Google Scholar] [CrossRef] [PubMed]
  66. Tong, L.I.; Saminathan, R.; Chang, C.W. Uncertainty Assessment of Non-Normal Emission Estimates Using Non-Parametric Bootstrap Confidence Intervals. J. Environ. Inform. 2016, 28, 61–70. [Google Scholar] [CrossRef]
  67. Anggani, N.L.; Amrullah, H.M.; Gemilang, D.S.A. Moran I Autocorrelation Study for Level Spatial Pattern Analysis. J. Indones. Sos. Teknol. 2023, 4, 1285–1291. [Google Scholar] [CrossRef]
  68. Zhang, S.; Li, J.; Zhang, T.; Feng, P.; Liu, W. Response of Vegetation to SPI and Driving Factors in Chinese Mainland. Agric. Water Manag. 2024, 291, 108625. [Google Scholar] [CrossRef]
  69. Zhu, Z.; Yang, Y.; Liu, B. Physics-Based Predictions of the Month-by-Month Summer Western North Pacific Anomalous Anticyclone. J. Clim. 2025, 38, 2187–2203. [Google Scholar] [CrossRef]
  70. Khansalari, S.; Majidi Dashli, O.; Nikzadfar, M.; Mollaarazi, A. Temporal and Spatial Changes of Dust in Golestan Province Using AOD (Aerosol Optical Depth) and the Affectability of This Province from the Deserts of Turkmenistan. J. Earth Space Phys. 2023, 49, 517–540. [Google Scholar] [CrossRef]
  71. Emadodin, I.; Reinsch, T.; Taube, F. Drought and Desertification in Iran. Hydrology 2019, 6, 66. [Google Scholar] [CrossRef]
  72. Zhao, F.; Zhang, M.; Zhu, S.; Zhang, X.; Ma, S.; Gao, Y.; Xia, J.; Wang, X.; Zhang, Y.; Zhang, S. Spatiotemporal Patterns of the Urban Thermal Environment and the Impact of Human Activities in Low-Latitude Plateau Cities. Int. J. Appl. Earth Obs. Geoinf. 2025, 142, 104703. [Google Scholar] [CrossRef]
  73. Wei, Z.; Miao, L.; Peng, J.; Zhao, T.; Meng, L.; Lu, H.; Peng, Z.; Cosh, M.H.; Fang, B.; Lakshmi, V. Bridging Spatio-Temporal Discontinuities in Global Soil Moisture Mapping by Coupling Physics in Deep Learning. Remote Sens. Environ. 2024, 313, 114371. [Google Scholar] [CrossRef]
  74. Yang, G.; Qiu, H.; Wang, N.; Yang, D.; Liu, Y. Tracking 35-Year Dynamics of Retrogressive Thaw Slumps across Permafrost Regions of the Tibetan Plateau. Remote Sens. Environ. 2025, 325, 114786. [Google Scholar] [CrossRef]
  75. Anees, S.A.; Mehmood, K.; Khan, W.R.; Shahzad, F.; Zhran, M.; Ayub, R.; Alarfaj, A.A.; Alharbi, S.A.; Liu, Q. Spatiotemporal Dynamics of Vegetation Cover: Integrative Machine Learning Analysis of Multispectral Imagery and Environmental Predictors. Earth Sci. Inf. 2025, 18, 152. [Google Scholar] [CrossRef]
  76. Mehmood, K.; Anees, S.A.; Rehman, A.; Pan, S.; Tariq, A.; Zubair, M.; Liu, Q.; Rabbi, F.; Khan, K.A.; Luo, M. Exploring Spatiotemporal Dynamics of NDVI and Climate-Driven Responses in Ecosystems: Insights for Sustainable Management and Climate Resilience. Ecol. Inf. 2024, 80, 102532. [Google Scholar] [CrossRef]
  77. Morsy, M.; Moursy, F.I.; Sayad, T.; Shaban, S. Climatological Study of SPEI Drought Index Using Observed and CRU Gridded Dataset over Ethiopia. Pure Appl. Geophys. 2022, 179, 3055–3073. [Google Scholar] [CrossRef]
  78. Nejadrekabi, M.; Eslamian, S.; Zareian, M.J. Spatial Statistics Techniques for SPEI and NDVI Drought Indices: A Case Study of Khuzestan Province. Int. J. Environ. Sci. Technol. 2022, 19, 6573–6594. [Google Scholar] [CrossRef]
  79. Bisht, H.; Shaloo; Suna, T.; Vishnoi, L.; Gautam, S.; Singh, D.K. Drought Assessment and Trend Analysis Using SPI and SPEI during Southwest Monsoon Season over Bundelkhand Region of Uttar Pradesh, India. Mausam 2023, 74, 119–128. [Google Scholar] [CrossRef]
  80. He, M.-Y.; Dong, J.-B.; Liu, X.; Kang, S.; Sun, Y.; Deng, L.; Zhang, N.; Zhang, X. Lithium Isotope Fractionation in Weinan Loess and Implications for Pedogenic Processes and Groundwater Impact. Glob. Planet. Change 2025, 252, 104865. [Google Scholar] [CrossRef]
  81. Yu, H.; Wang, L.; Zhang, J.; Chen, Y. A Global Drought-Aridity Index: The Spatiotemporal Standardized Precipitation Evapotranspiration Index. Ecol. Indic. 2023, 153, 110484. [Google Scholar] [CrossRef]
  82. Wang, M.; He, G.; Hu, T.; Yang, M.; Zhang, Z.; Zhang, Z.; Wang, G.; Li, H.; Gao, W.; Liu, X. Innovative Hybrid Algorithm for Simultaneous Land Surface Temperature and Emissivity Retrieval: Case Study with SDGSAT-1 Data. Remote Sens. Environ. 2024, 315, 114449. [Google Scholar] [CrossRef]
  83. Zhang, Y.; Wu, X. Global Space-Time Patterns of Sub-Daily Extreme Precipitation and Its Relationship with Temperature and Wind Speed. Environ. Res. Lett. 2025, 20, 084019. [Google Scholar] [CrossRef]
  84. Liu, T.; Yu, L.; Yan, Z.; Li, X.; Bu, K.; Yang, J. Enhanced Climate Mitigation Feedbacks by Wetland Vegetation in Semi-arid Compared to Humid Regions. Geophys. Res. Lett. 2025, 52, e2025GL115242. [Google Scholar] [CrossRef]
  85. Zhu, Z.; Shao, L.; Lu, R.; Hua, W. Two Contrasting Tropical Convection Modes from the Eastern Pacific to Northern Africa That Drive Eurasian Teleconnections in Boreal Summer. npj Clim. Atmos. Sci. 2025, 8, 56. [Google Scholar] [CrossRef]
  86. Jiang, C.; Wang, Y.; Yang, Z.; Zhao, Y. Do Adaptive Policy Adjustments Deliver Ecosystem-Agriculture-Economy Co-Benefits in Land Degradation Neutrality Efforts? Evidence from Southeast Coast of China. Env. Monit. Assess. 2023, 195, 1215. [Google Scholar] [CrossRef]
  87. Moges, A.; Tegbaru, G. Estimation of Climate and Watershed Management-Induced Potential Soil Loss Using the USLE in Demjer Watershed, Blue Nile Basin, Ethiopia. Geol. Ecol. Landsc. 2024, 1–19. [Google Scholar] [CrossRef]
  88. Mehmood, K.; Anees, S.A.; Rehman, A.; Tariq, A.; Liu, Q.; Muhammad, S.; Rabbi, F.; Pan, S.; Hatamleh, W.A. Assessing Forest Cover Changes and Fragmentation in the Himalayan Temperate Region: Implications for Forest Conservation and Management. J. For. Res. 2024, 35, 82. [Google Scholar] [CrossRef]
  89. Ding, Y.; Zheng, X.; Zhao, K.; Xin, X.; Liu, H. Quantifying the Impact of NDVIsoil Determination Methods and NDVIsoil Variability on the Estimation of Fractional Vegetation Cover in Northeast China. Remote Sens. 2016, 8, 29. [Google Scholar] [CrossRef]
  90. Mirzabaev, A.; Ahmed, M.; Werner, J.; Pender, J.; Louhaichi, M. Rangelands of Central Asia: Challenges and Opportunities. J. Arid Land 2016, 8, 93–108. [Google Scholar] [CrossRef]
  91. Kumar, B.P.; Babu, K.R.; Anusha, B.N.; Rajasekhar, M. Geo-Environmental Monitoring and Assessment of Land Degradation and Desertification in the Semi-Arid Regions Using Landsat 8 OLI/TIRS, LST, and NDVI Approach. Environ. Chall. 2022, 8, 100578. [Google Scholar] [CrossRef]
  92. Chang, S.; Chen, H.; Wu, B.; Nasanbat, E.; Yan, N.; Davdai, B. A Practical Satellite-Derived Vegetation Drought Index for Arid and Semi-Arid Grassland Drought Monitoring. Remote Sens. 2021, 13, 414. [Google Scholar] [CrossRef]
  93. Beguería, S.; Peña-Angulo, D.; Trullenque-Blanco, V.; González-Hidalgo, C. MOPREDAScentury: A Long-Term Monthly Precipitation Grid for the Spanish Mainland. Earth Syst. Sci. Data 2023, 15, 2547–2575. [Google Scholar] [CrossRef]
  94. Beguería, S.; Vicente-Serrano, S.M. SPEI: Calculation of the Standardized Precipitation–Evapotranspiration Index (R Package Version 1.8.1); Comprehensive R Archive Network (CRAN), R Foundation for Statistical Computing: Vienna, Austria, 2023; Available online: https://CRAN.R-project.org/package=SPEI (accessed on 28 June 2025).
  95. Pontius, R.G., Jr.; Millones, M. Problems and Solutions for Kappa-Based Indices of Agreement. Stud. Model. Sense Mak. Planet Earth 2008, 8. [Google Scholar]
  96. Sun, L.; Feng, Z.; Shao, Y.; Wang, L.; Su, J.; Ma, T.; Lu, D.; An, J.; Pang, Y.; Fahad, S.; et al. The development of a set of novel low cost and data processing-free measuring instruments for tree diameter at breast height and tree position. Forests 2023, 14, 891. [Google Scholar] [CrossRef]
  97. Allington, G.; Kreitzer, N. Detecting Land Cover Change in Rangelands. 2023. Available online: https://google-earth-engine.com/Terrestrial-Applications-part-2/Detecting-Land-Cover-Change-in-Rangelands/ (accessed on 28 June 2025).
  98. Bonney, M.T.; He, Y.; Myint, S.W. Contextualizing the 2019–2020 Kangaroo Island Bushfires: Quantifying Landscape-Level Influences on Past Severity and Recovery with Landsat and Google Earth Engine. Remote Sens. 2020, 12, 3942. [Google Scholar] [CrossRef]
  99. Kennedy, R.E.; Yang, Z.; Cohen, W.B. Detecting trends in forest disturbance and recovery using yearly Landsat time series: 1. LandTrendr—Temporal segmentation algorithms. Remote Sens. Environ. 2010, 114, 2897–2910. [Google Scholar] [CrossRef]
Figure 1. Study area in central Iraq. Satellite basemap with governorate boundaries shown using semi-transparent polygons. The study area (Babil and Al-Qadisiyah) is emphasised with diagonal hatching and a solid red outline. The inset locates the two governorates within the national context and shows broad elevation variation (m).
Figure 1. Study area in central Iraq. Satellite basemap with governorate boundaries shown using semi-transparent polygons. The study area (Babil and Al-Qadisiyah) is emphasised with diagonal hatching and a solid red outline. The inset locates the two governorates within the national context and shows broad elevation variation (m).
Remotesensing 17 03343 g001
Figure 2. Analytical workflow for degradation risk mapping and driver attribution.
Figure 2. Analytical workflow for degradation risk mapping and driver attribution.
Remotesensing 17 03343 g002
Figure 3. LULC classification maps of the study area for the years (A) 2000, (B) 2007, (C) 2015, and (D) 2023. The maps show spatial distribution and temporal changes across five LULC classes: Cultivated Agriculture (CA), Herbaceous/Orchards (HO), Bare Land (BL), Urban Land (UL), and Water Bodies (WB).
Figure 3. LULC classification maps of the study area for the years (A) 2000, (B) 2007, (C) 2015, and (D) 2023. The maps show spatial distribution and temporal changes across five LULC classes: Cultivated Agriculture (CA), Herbaceous/Orchards (HO), Bare Land (BL), Urban Land (UL), and Water Bodies (WB).
Remotesensing 17 03343 g003
Figure 4. Class-wise net change in LULC area (km2) across four-time intervals. Positive values indicate net expansion, while negative values denote net contraction. CA = Cultivated Agriculture, HO = Herbaceous/Orchards, WB = Water/Waterlogged Bodies, UL = Urban Land, BL = Bare Land.
Figure 4. Class-wise net change in LULC area (km2) across four-time intervals. Positive values indicate net expansion, while negative values denote net contraction. CA = Cultivated Agriculture, HO = Herbaceous/Orchards, WB = Water/Waterlogged Bodies, UL = Urban Land, BL = Bare Land.
Remotesensing 17 03343 g004
Figure 5. Spatial patterns of long-term vegetation trends based on Fractional Vegetation Cover (FVC) from 2000 to 2023 across the Babil and Al-Qadisiyah governorates, Iraq. (A) Classified FVC trends categorized into five classes: Very High (≤–10 pp/decade), High (–10 to –5 pp/decade), Moderate (–5 to +5 pp/decade), Low (+5 to +10 pp/decade), and Very Low (≥ +10 pp/decade). (B) Continuous map of statistically significant FVC trends (p < 0.05), showing Sen’s Slope estimates in percentage points per decade. Note: pp = percentage points. Negative values indicate vegetation decline; positive values represent vegetation recovery.
Figure 5. Spatial patterns of long-term vegetation trends based on Fractional Vegetation Cover (FVC) from 2000 to 2023 across the Babil and Al-Qadisiyah governorates, Iraq. (A) Classified FVC trends categorized into five classes: Very High (≤–10 pp/decade), High (–10 to –5 pp/decade), Moderate (–5 to +5 pp/decade), Low (+5 to +10 pp/decade), and Very Low (≥ +10 pp/decade). (B) Continuous map of statistically significant FVC trends (p < 0.05), showing Sen’s Slope estimates in percentage points per decade. Note: pp = percentage points. Negative values indicate vegetation decline; positive values represent vegetation recovery.
Remotesensing 17 03343 g005
Figure 6. Spatial distribution of vegetation degradation severity and timing across the study area. (A) Year-of-Disturbance (YoD) stratified into four temporal intervals (1990s, 2000s, 2010s, 2020s), illustrating the decadal emergence of degradation hotspots. (B) Composite degradation severity classified as Low, Moderate, and High based on a normalized index combining spectral magnitude, rate of degradation, and signal-to-noise ratio (DSNR).
Figure 6. Spatial distribution of vegetation degradation severity and timing across the study area. (A) Year-of-Disturbance (YoD) stratified into four temporal intervals (1990s, 2000s, 2010s, 2020s), illustrating the decadal emergence of degradation hotspots. (B) Composite degradation severity classified as Low, Moderate, and High based on a normalized index combining spectral magnitude, rate of degradation, and signal-to-noise ratio (DSNR).
Remotesensing 17 03343 g006
Figure 7. Spatial distribution of severe climatic drought and joint ecological stress in the study area. Panels (AC) show the frequency of severe drought (SPEI < –1.5) at 3-, 6-, and 12-month accumulation scales. Panels (DF) depict drought–vegetation stress co-occurrence (SPEI + VHI < 40) without lag, while panels (GI) show the same with a one-month lag applied to VHI. Warmer colors represent higher frequencies of stress, highlighting spatial variability and the influence of temporal scale and ecological lag on drought impact detection.
Figure 7. Spatial distribution of severe climatic drought and joint ecological stress in the study area. Panels (AC) show the frequency of severe drought (SPEI < –1.5) at 3-, 6-, and 12-month accumulation scales. Panels (DF) depict drought–vegetation stress co-occurrence (SPEI + VHI < 40) without lag, while panels (GI) show the same with a one-month lag applied to VHI. Warmer colors represent higher frequencies of stress, highlighting spatial variability and the influence of temporal scale and ecological lag on drought impact detection.
Remotesensing 17 03343 g007
Figure 8. Spatial prediction outputs from the XGBoost degradation classification model. Continuous probability map of vegetation degradation, with higher values (red) indicating greater model-estimated disturbance likelihood. Discrete degradation risk classification based on probability thresholds: Low (<0.4), Moderate (0.4–0.7), and High (>0.7). High-risk zones are spatially concentrated in the southern and southwestern sectors, reflecting localized vulnerability under cumulative climatic and anthropogenic stressors.
Figure 8. Spatial prediction outputs from the XGBoost degradation classification model. Continuous probability map of vegetation degradation, with higher values (red) indicating greater model-estimated disturbance likelihood. Discrete degradation risk classification based on probability thresholds: Low (<0.4), Moderate (0.4–0.7), and High (>0.7). High-risk zones are spatially concentrated in the southern and southwestern sectors, reflecting localized vulnerability under cumulative climatic and anthropogenic stressors.
Remotesensing 17 03343 g008
Figure 9. SHAP-based interpretability of the XGBoost degradation classification model. (A) Ranked importance of predictors based on mean absolute SHAP values, indicating relative contributions of climatic, land-use, and spatial drivers to predicted degradation probability. (B) Top ten pairwise SHAP interaction strengths, illustrating key synergies particularly among drought indices, agricultural intensity, and hydrological proximity that influence model predictions beyond additive effects.
Figure 9. SHAP-based interpretability of the XGBoost degradation classification model. (A) Ranked importance of predictors based on mean absolute SHAP values, indicating relative contributions of climatic, land-use, and spatial drivers to predicted degradation probability. (B) Top ten pairwise SHAP interaction strengths, illustrating key synergies particularly among drought indices, agricultural intensity, and hydrological proximity that influence model predictions beyond additive effects.
Remotesensing 17 03343 g009
Table 1. Datasets employed in this study.
Table 1. Datasets employed in this study.
Dataset/Derived ProductPrimary SourceNative/Working ResolutionTemporal Coverage
Landsat imagery and derived indices (NDVI, FVC, NBR, LST, VHI)GEE—USGS Collection 230 m2000–2023
SPEI (monthly; 1-, 3-, 6-, 12-month)TerraClimate4 km → 30 m *2000–2023
Joint ecological stress (SPEI–VHI co-occurrence, 1-month lag)(TerraClimate + Landsat)30 m2000–2023
Annual LULC classificationsLandsat (RF)30 m2000–2023
Distance to roads, rivers, waterbodiesOpenStreetMap10 m → 30 m *Current (static)
Population densityWorldPop100 m → 30 m *2000–2023
Agricultural-intensity metricsProvincial agricultural statisticsDistrict polygons → 30 m *2000–2023
Urban expansion rateGHSL built-up change30 m2000–2023
* Resampled via bilinear (continuous) or nearest-neighbour (categorical) interpolation to align with the Landsat grid.
Table 2. Sen’s slope estimator was used to quantify annual vegetation change, expressed as percentage change per decade.
Table 2. Sen’s slope estimator was used to quantify annual vegetation change, expressed as percentage change per decade.
Slope Range (%/Decade)Trend Category
<–10Severe degradation
–10 to –5Moderate degradation
–5 to +5Stable
+5 to +10Moderate recovery
>+10Strong recovery
Table 3. LULC transition matrix from 2000 to 2023 with row and column totals (km2). Rows represent the origin class in 2000, and columns show the corresponding class in 2023. The row totals represent the total area per class in 2000, while the column totals show the resulting area per class in 2023.
Table 3. LULC transition matrix from 2000 to 2023 with row and column totals (km2). Rows represent the origin class in 2000, and columns show the corresponding class in 2023. The row totals represent the total area per class in 2000, while the column totals show the resulting area per class in 2023.
CAHOWBULBLRow Total
CA8746.5267.7842.7967.81429.029553.9
HO197.22623.4711.8912.1511.862856.57
WB8.096.45221.570.162.74239.01
UL69.0111.010.15394.493.83478.49
BL356.5823.3673.0612.272420.252885.52
Column Total9377.372932.08349.46486.882867.716,013.48
Table 4. Summary of XGBoost classification performance metrics for the training and spatially independent test sets. The model exhibited strong generalization with balanced sensitivity and specificity across classes, achieving an overall accuracy of 79.2% and AUC of 0.884 on the test set. Performance was slightly higher on the training set (accuracy = 83.5%, AUC = 0.925), indicating minimal overfitting. Consistent F1-scores and precision for the degraded class confirm the model’s reliability in detecting vegetation degradation.
Table 4. Summary of XGBoost classification performance metrics for the training and spatially independent test sets. The model exhibited strong generalization with balanced sensitivity and specificity across classes, achieving an overall accuracy of 79.2% and AUC of 0.884 on the test set. Performance was slightly higher on the training set (accuracy = 83.5%, AUC = 0.925), indicating minimal overfitting. Consistent F1-scores and precision for the degraded class confirm the model’s reliability in detecting vegetation degradation.
MetricTrainingTest
Accuracy83.5%79.2%
Kappa0.66970.5839
Sensitivity (Degraded)0.79730.7532
Specificity (Stable)0.87240.8307
Precision (PPV)0.81650.8165
F1 Score0.78350.7835
AUC0.9250.884
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Al-Tameemi, N.; Xuexia, Z.; Shahzad, F.; Mehmood, K.; Linying, X.; Zhou, J. From Trends to Drivers: Vegetation Degradation and Land-Use Change in Babil and Al-Qadisiyah, Iraq (2000–2023). Remote Sens. 2025, 17, 3343. https://doi.org/10.3390/rs17193343

AMA Style

Al-Tameemi N, Xuexia Z, Shahzad F, Mehmood K, Linying X, Zhou J. From Trends to Drivers: Vegetation Degradation and Land-Use Change in Babil and Al-Qadisiyah, Iraq (2000–2023). Remote Sensing. 2025; 17(19):3343. https://doi.org/10.3390/rs17193343

Chicago/Turabian Style

Al-Tameemi, Nawar, Zhang Xuexia, Fahad Shahzad, Kaleem Mehmood, Xiao Linying, and Jinxing Zhou. 2025. "From Trends to Drivers: Vegetation Degradation and Land-Use Change in Babil and Al-Qadisiyah, Iraq (2000–2023)" Remote Sensing 17, no. 19: 3343. https://doi.org/10.3390/rs17193343

APA Style

Al-Tameemi, N., Xuexia, Z., Shahzad, F., Mehmood, K., Linying, X., & Zhou, J. (2025). From Trends to Drivers: Vegetation Degradation and Land-Use Change in Babil and Al-Qadisiyah, Iraq (2000–2023). Remote Sensing, 17(19), 3343. https://doi.org/10.3390/rs17193343

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop