Tracking Quarter-Century Spatio-Temporal Soil Salinization Dynamics in Semi-Arid Landscapes Using Earth Observation and Machine Learning

Highlights

What are the main findings?

• A 25-year Remote Sensing–Machine Learning (RS–ML) analysis reveals long-term persistent hotspots (≥32 dS/m) and a general shift from low to moderate soil salinity.
• SVR outperformed GBT and RF models, achieving the highest predictive accuracy and thus was selected for the spatiotemporal soil-salinity modeling. Given the inherent noise and complexity of archival RS datasets, performances remain within acceptable ranges, reflecting stable generalization.

What are the implications of the main findings?

• Climatic and topographic factors significantly improved model performance and proved essential for understanding salinity dynamics controlled by episodic inundation, shallow saline groundwater, and hydro-climatic variability.
• Over time, moderately to highly saline areas (≥16 dS/m) expanded by approximately 10%, driven by recurrent droughts, intense evaporation, and inefficient drainage.

Abstract

Soil salinization represents a critical constraint to sustainable agriculture in arid and semi-arid regions, where salinity threatens soil productivity, water quality, and ecosystem resilience. Soil salinity pattern prediction is complicated by tightly coupled landscape hydro-climatic processes, wherein the central Sabkha acts as a persistent salt sink, episodic inundation and intense evaporation concentrate dissolved salts, and a shallow saline groundwater table interacts with the semi-arid climate to drive surface salinization. Conventional mapping is laborious and lacks the precision needed to capture the spatio-temporal dynamics of soil salinity across landscapes. This study developed an integrated framework uniting multi-temporal Landsat imagery (2000–2025), hypsometric data, climatic indicators, and in situ soil electrical conductivity (ECe) measurements to model soil salinity dynamics using machine learning (ML), over the Sehb El Masjoune (SEM) semi-arid region, Morocco. A total of 233 soil samples were collected in the investigated area in 2022, 2023, 2024, and 2025 to assess the spatial variability to calibrate and validate modeling findings. To this end, three predictive algorithms, i.e., Gradient-Boosted Trees (GBT), Support Vector Regression (SVR), and Random Forest (RF) were assessed. Our findings showed that SVR achieved the highest predictive capability (R² = 0.76; RMSE = 32.91 dS/m), whereas SVR-based salinity maps revealed a distinct spatial organization of salinization processes, characterized by extremely saline soils (≥64 dS/m) concentrated in the central study area (i.e., SEM center) and a progressive decline toward adjacent agricultural lands (0–8 dS/m). Our results demonstrated that from 2000 to 2025, moderately to highly saline areas (≥16 dS/m) expanded by nearly 10%, driven by recurrent droughts and inefficient drainage. Hydroclimatic analysis confirmed that dry years (SPI: Standardized Precipitation Index ≤ −0.5) promoted net salinity build-up through the expansion and persistence of moderate-to-high salinity classes (≥16 dS/m), whereas wet years (SPI ≥ +0.5) favored temporary leaching and partial recovery, mainly within the low-to-moderate range. This integrative remote sensing–ML approach provides a robust and scalable framework for operational soil salinity monitoring, offering valuable insights for sustainable land-use planning in similar Sabkha’s data-scarce agroecosystems.

1. Introduction

Salinity of soil is a growing challenge threatening agricultural productivity, soil health, and food security, especially in dry climates [1,2]. This form of land degradation is often the result of complex interactions among natural processes such as evapotranspiration, topographic concentration of runoff, and shallow saline groundwater and anthropogenic pressures, including excessive irrigation, inefficient drainage systems, and inappropriate land management practices [3,4]. In Morocco, salinization is a widespread issue affecting several productive plains, including the Sehb El Masjoune area, where historical irrigation practices combined with endorheic geomorphology and high evaporation rates create favorable conditions for salt accumulation at the surface and in root zones [5,6]. As drought cycles have intensified and become more frequent in recent years, Morocco has faced escalating challenges with salt-affected area, where shifting climate patterns are accelerating soil degradation and expanding the area of compromised agricultural productivity [7,8]. To effectively address this issue and plan for long-term sustainability, it is imperative to develop accurate, wide-area maps that identify saline soils.

The dynamic and multifaceted processes of soil salinization make predicting its temporal occurrence over large expanses particularly difficult. Until recently, geostatistical models (i.e., through digital soil mapping (DSM) techniques) were used with GISs in many previous studies, to assess, map, and monitor saline lands. However, their prediction accuracy was affected by the number of samples and their spatial distribution over the area [9]. Compounding the challenge, controlling for the spread of soil salinity requires rigorous field and laboratory analyses. Consequently, conventional mapping remains an expensive and laborious procedure, ill-suited for frequent, large-scale monitoring. These severe constraints necessitate more practical alternatives, especially for regions with limited technical capacity.

Furthermore, remote sensing (RS) data can facilitate the creation of large-scale, highly accurate DSM by offering rapid data acquisition, broad spatial coverage, and multi-temporal monitoring capabilities [10,11]. Indeed, RS-based monitoring can be broadly categorized into direct assessments of bare soil reflectance and indirect assessments [12,13]. For instance, a growing body of research has demonstrated strong correlations between vegetation indices and soil salinity, highlighting their utility as indirect indicators of salinity levels [12]. Moreover, the use of salinity-specific spectral indices and composite indices has expanded considerably, improving the accuracy of salinity mapping [11,14]. In addition to spectral information, incorporating covariates such as elevation, aspect, and slope has shown strong effect in enhancing the discrimination of salinized soils [13,15]. Many recent studies have increasingly combined in situ data with spectral indices and topographical covariates to build predictive models that accurately map soil salinity using diverse satellite platforms [1,11,16,17].

Despite its severity, limited research has investigated the spatiotemporal dynamics of salinization in Morocco, and specifically in the Sehb El Masjoune area, using modern geospatial techniques. Traditional mapping approaches, often based on kriging or sparse field surveys, fail to capture the continuous and dynamic nature of soil salinity distribution across time and space [18]. Moreover, many existing assessments lack the ability to explain the underlying environmental drivers that contribute to soil salinity variation. This creates a significant knowledge gap, particularly in Morocco, where integrated soil, climate, and remote sensing frameworks have not yet been operationalized for long-term salinity surveillance.

Machine learning assessments of soil salinity generally produce spatial patterns of classification labels (ranging from non-saline to highly saline) that correspond to landscape features like low-lying, poorly drained fields, irrigation boundaries, and areas with shallow groundwater, while nearby uplands are typically less impacted, trends that are consistently observed across remote sensing-based geostatistical methods and data-driven models [19,20]. Machine learning algorithms, e.g., Gradient-Boosted Trees (GBT), Support Vector Regression (SVR), and Random Forest (RF), are increasingly employed for soil salinity prediction, due to their ability to model complex nonlinear relationships between remote sensing variables and measured electrical conductivity (EC) [21,22,23,24]. Findings from the Gonbad drainage area (situated in the northern part of Iran) indicate that, when provided with suitable spectral and environmental indices, these models can accurately replicate surface salinity distributions [23]. Likewise, studies conducted at multiple depths in the Yellow River Delta reveal that SVR and RF models effectively reconstruct vertical salinity profiles using thermal, moisture, vegetation, and salinity indices, further validating their effectiveness for salinity prediction [24].

In comparative analyses, machine learning algorithms deliver precise class predictions and can illustrate model variation through ensemble differences, while Bayesian hierarchical models systematically convey uncertainty to pixel-level credible intervals, depth-integrated assessments, and exceedance probability maps that pinpoint locations where soil salinity may surpass management limits [25,26]. Uncertainty surfaces created alongside class maps, such as coefficients of variation, standard deviation, credible interval widths, or probability of exceedance layers, consistently identify data scarce or environmentally unusual regions and are likely to increase with depth as observations diminish, reflecting broader digital soil mapping insights regarding profile-level uncertainty [25,27]. Integrating spatial predictions with these uncertainty metrics enhances decision-making by prioritizing actions (e.g., targeted leaching, drainage, or monitoring) in areas where both predicted salinity class and model reliability are high, while identifying uncertain hotspots for additional sampling or improved model inputs [19,26].

A comprehensive temporal evaluation of salinization as influenced by drought should initially categorize hydroclimatic conditions utilizing the Standardized Precipitation Index (SPI) across various temporal scales (e.g., 3, 6, and 12 months) to differentiate agricultural drought from hydrological drought and to synchronize deficit timelines with soil and groundwater responses [28]. Empirical investigations within arid plains reveal that the intensity of drought correlates with surface soil salinity: for instance, dry years as defined by the SPI were associated with markedly elevated soil electrical conductivity, with a singular inverse year interpreted as a delayed system reaction [7,29]. In coastal aquifers, meteorological drought diminishes recharge and lowers groundwater levels, while increased extraction to satisfy irrigation needs amplifies stress; even subsequent to groundwater level recovery, water quality can remain compromised (notably increased chloride levels), indicating a hysteretic relationship between water quantity and salinity during and following severe drought events [28]. Furthermore, drought-induced soil salinization is evident in wetlands, where prolonged drought leads to salinity spikes that disrupt ecosystem functioning [30]. Analyses conducted on a seasonal basis additionally reveal redistributions among classes of salinity risk rather than a linear progression [19].

From a practical standpoint, the integration of SPI series with probability-focused mapping of drought frequency and severity (e.g., the Probability of Drought Severity, PDS, method) facilitates a spatially explicit monitoring of when and where salinity pressures are anticipated to escalate [31]. Collectively, these insights advocate for a methodological framework that aligns drought occurrence with SPI, analyzes salinity responses while considering delays and hysteresis, and differentiates by geomorphic setting (endorheic versus coastal) to project the extent and duration of salinization associated with drought conditions [16,21,24]. This approach combines field data and satellite imagery to accurately monitor soil salinity over large areas.

Accordingly the objectives of this study are (i) to evaluate the overall accuracy and the performance of machine learning models for soil salinity estimation, together with error estimates, (ii) to generate spatiotemporal maps of soil salinity, by integrating SPI-based hydroclimatic regimes (wet/normal/dry), and ultimately (iii) to measure the change in saline lands over time at the scale over a quarter-century time period (2000–2005), to support sustainable land and water management.

2. Materials and Methods

2.1. Study Area

The study area is located in the semi-arid-to-arid Bahira Plain of central Morocco, and more specifically in the Sehb El Masjoune region, situated approximately 60 km north of Marrakech [5,32,33], and extends between latitudes 32.009° to 32.194°N and longitudes 7.979° to 7.528°W (Figure 1). This region is characterized by a central topographic depression (i.e., seasonal dry lake), covering an area of about 32 km², which periodically transforms into a wetland during heavy rains before rapidly evaporating, leaving behind salt deposits that contribute to soil salinization [5,6]. The surrounding landscape includes the Gantour Plateau to the north and the Jbilets Mountains to the south, which influence the hydrogeological dynamics of the area [6].

The Sehb El Masjoune sabkha is climatically semi-arid, with short, mild winters and long, hot summers. The average annual precipitation is less than 200 mm, while evaporation rates are high, reaching up to 2700 mm annually [5,32]. These climatic conditions, combined principally with poor drainage and saline irrigation water use, have led to widespread soil salinity, affecting agricultural productivity and leading to the abandonment of farmland in some areas [5,34]. The region’s soils are predominantly fine-to-medium-textured, with high electrical conductivity (ECe) values ranging from 0.5 to 235 dS/m, classifying them as saline to extremely saline [32,34].

Agriculture in the area is primarily traditional, with crops such as barley, wheat, and olive trees being cultivated in less saline zones [32]. However, soil salinity has significantly reduced yields, particularly for sensitive crops, making the reclamation of salt-affected soils a critical priority [34]. The region is also characterized by the existence of halophyte vegetation, such as Anabasis articulata and Salsola soda, which thrive in saline conditions and are indicators of soil salinity [5,32].

Hydrogeologically, the area is part of the Bahira plain, a closed basin where groundwater converges toward the central Sabkha, leading to shallow piezometric levels and high salinity in aquifer systems [6]. The groundwater is brackish to saline, with chloride-dominated facies, and its salinization is attributed to evaporation, dissolution of geological formations, and anthropogenic activities such as irrigation with saline water [6].

The methodology combines multi-source geospatial data (i.e., SRTM, Landsat) and machine learning (i.e., Gradient-Boosted Trees (GBT), Support Vector Regression (SVR), and Random Forest (RF)) to map soil salinity over time. We integrated satellite imagery, field measurements, climate records, and topographic data to build predictive models of soil salinity, which were then applied to produce yearly salinity maps. The flowchart in Figure 2 offers a summary of the different methodological phases (i.e., data pre-processing, feature extraction, model training/validation, and spatiotemporal analysis).

The methodological framework integrates multi-source geospatial datasets and machine learning techniques to model and map soil salinity dynamics. As shown in Figure 2, the workflow begins with the acquisition and preprocessing of field soil salinity measurements, Landsat spectral indices and thermal bands, MODIS LST resampled to 30 m, SRTM-derived topographic variables (elevation, slope, aspect), and station-based precipitation records used to compute the Standardized Precipitation Index (SPI). Remote sensing-derived indices and predictors are harmonized and stacked to create the unified feature set used for model development, followed by correlation analysis to identify relationships between soil samples and remote sensing predictors. The resulting dataset is partitioned into training (65%), validation (25%), and independent testing (10%) subsets. Three machine learning algorithms, Gradient-Boosted Trees (GBT), Support Vector Regression (SVR), and Random Forests (RF), are then trained and evaluated using regression metrics (R², RMSE, MAE, bias). The best-performing model is subsequently applied to generate yearly soil salinity maps for 2000–2025, from which spatial probability maps, temporal trends, and feature importance analyses are derived. This structured overview encapsulates the key phases of data preprocessing, feature extraction, model development, and spatio-temporal analysis.

2.2. Soil Sampling and Analysis

During the summer of 2022, 2023, 2024, and 2025, four soil sampling campaigns were conducted in the Sehb El Masjoune area to assess the spatial and temporal variability of soil salinity. To this end, a total of 233 soil samples were collected along a north–south gradient (Figure 1). At each sampling location, a 30 m² quadrat was established, from which five topsoil subsamples (0–15 cm depth), were collected using a soil auger, one from each corner and one from the center. These subsamples were homogenized to form a single composite sample, ensuring a representative characterization of surface salinity variation. The 2022 dataset contained the largest number of samples obtained from bare saline soils (131 soil samples), making it the most suitable and statistically robust dataset for model calibration.

Soil samples were oven-dried at 38 °C for 24 h. After drying, plant roots and residues were excluded, and ground to pass through a 2 mm sieve. Fractions of soil texture were determined using a hydrometer and classified according to the USDA textural triangle classification.

Soil salinity was determined based on the ECe (dS/m). The saturated paste extract method (USSL, 1954) [35] was primarily used, whereby saturated pastes were prepared from air-dried and sieved soil samples, equilibrated for 24 h, and the extracts were obtained by vacuum filtration and analyzed with a calibrated conductivity meter (SevenDirect SD23, Mettler-Toledo, Greifensee, Switzerland). For the 2022 sampling campaign, a subset of samples was analyzed using the saturated paste extract method, while others were measured using the soil-to-water (1:5) extract ratio (EC1:5) to optimize analytical throughput. The EC1:5 extracts were prepared following the procedure of Ouzemou et al. [32], by mixing 20 g of soil with 100 mL of distilled water, shaking for 30 min, filtering through a Whatman No. 2 cellulose filter, and measuring EC within 3 min. To ensure data consistency, the EC1:5 values were converted to their equivalent ECe using the polynomial regression relationship established for semi-arid Moroccan soils (R² = 0.98) as described by Ouzemou et al. [32]. Thus, all salinity data were standardized and reported as ECe (dS/m). For subsequent years (2023–2025), all samples were analyzed exclusively using the saturated paste extract method to maintain analytical uniformity and accuracy across the multi-year dataset.

All models were trained and validated using this summer 2022 field EC measurement subset and subsequently tested with independent test datasets from summer 2022, 2023, 2024, and 2025 (Figure 1). Following validation, the trained models were applied for retrospective prediction across the entire climatical and optical (Landsat) archive data (2000–2025), to maintain temporal and phenological consistency in the spatio-temporal salinity analysis.

2.3. Remote Sensing Data and Image Processing

2.3.1. Multitemporal Landsat Datasets

In this study, time-series salinity mapping was performed using multitemporal Landsat satellite imagery over different periods (

Table 1. Summary of the Landsat scenes utilized and their main characteristics. Abbreviations: TM (Thematic Mapper); ETM+ (Enhanced Thematic Mapper Plus); OLI/TIRS (Operational Land Imager/Thermal Infrared Sensor).

Period	Satellite	Sensor	Spatial Resolution	Scene IDs
2000	Landsat 5	TM	30 m	LANDSAT/LT05/C02/T1_L2/LT05_202038_20000704
2001				LANDSAT/LT05/C02/T1_L2/LT05_202038_20010723
2002				LANDSAT/LT05/C02/T1_L2/LT05_202038_20020608
2003				LANDSAT/LT05/C02/T1_L2/LT05_202038_20030713
2004				LANDSAT/LT05/C02/T1_L2/LT05_202038_20040715
2005				LANDSAT/LT05/C02/T1_L2/LT05_202038_20050718
2006				LANDSAT/LT05/C02/T1_L2/LT05_202038_20060721
2007				LANDSAT/LT05/C02/T1_L2/LT05_202038_20070708
2008				LANDSAT/LT05/C02/T1_L2/LT05_202038_20080624
2009				LANDSAT/LT05/C02/T1_L2/LT05_202038_20090713
2010				LANDSAT/LT05/C02/T1_L2/LT05_202038_20100716
2011				LANDSAT/LT05/C02/T1_L2/LT05_202038_20110804
2012	Landsat 7	ETM+	30 m	LANDSAT/LE07/C02/T1_L2/LE07_202038_20120611
2013	Landsat 8	OLI/TIRS	30 m (with TIRS resampled to 30)	LANDSAT/LC08/C02/T1_L2/LC08_202038_20130724
2014				LANDSAT/LC08/C02/T1_L2/LC08_202038_20140727
2015				LANDSAT/LC08/C02/T1_L2/LC08_202038_20150714
2016				LANDSAT/LC08/C02/T1_L2/LC08_202038_20160716
2017				LANDSAT/LC08/C02/T1_L2/LC08_202038_20170703
2018				LANDSAT/LC08/C02/T1_L2/LC08_202038_20180722
2019				LANDSAT/LC08/C02/T1_L2/LC08_202038_20190709
2020				LANDSAT/LC08/C02/T1_L2/LC08_202038_20200711
2021				LANDSAT/LC08/C02/T1_L2/LC08_202038_20210714
2022				LANDSAT/LC08/C02/T1_L2/LC08_202038_20220701
2023				LANDSAT/LC08/C02/T1_L2/LC08_202038_20230720
2024				LANDSAT/LC08/C02/T1_L2/LC08_202038_20240706
2025				LANDSAT/LC08/C02/T1_L2/LC08_202038_20250725

). To this end, Landsat 5 Thematic Mapper (TM) scenes covering the period from 2000 to 2011 were used. Landsat 5 TM data were exclusively employed for this period to ensure a homogeneous and radiometrically consistent time series, minimizing cross-sensor variability and avoiding the scan line corrector (SLC)-related artifacts associated with Landsat 7 (ETM+). Landsat 7 Enhanced Thematic Mapper Plus (ETM+) data for 2012 and Landsat 8 Operational Land Imager (OLI) imagery spanning 2013–2025 were used to capture long-term variations in soil surface conditions (Table 1). Surface reflectance values from key spectral bands, particularly blue, green, red, near-infrared (NIR), and shortwave infrared (SWIR), were extracted to derive salinity-sensitive spectral indices. These indices were subsequently employed as input variables for machine learning models used in the soil salinity inversion (Refer to Section 2.4.2). By leveraging the continuity and spatial consistency of Landsat observations, this approach facilitated the monitoring of salt-affected soils across multiple years and growing seasons, supporting a robust spatiotemporal analysis of soil salinization dynamics.

2.3.2. Digital Elevation Model Data

A 30 m resolution Digital Elevation Model (DEM) based on Shuttle Radar Topography Mission (SRTM) data was used to obtain topographic parameters: altitude, gradient, and directional orientation. These metrics are essential in controlling the spatial distribution of soil salinity by influencing surface water movement, drainage efficiency, and salt accumulation in low-lying or poorly drained areas [24,36]. Elevation affects both groundwater depth and capillary rise potential, whereas slope governs overland flow and infiltration rates, which are critical to salt leaching and redistribution processes [11]. Aspect can further modulate soil moisture and evaporation patterns by altering exposure to solar radiation, especially in arid and semi-arid regions [37]. Incorporating these topographic variables into salinity prediction models enhances their physical interpretability and improves spatial accuracy, especially when used alongside spectral and vegetation-derived indices [18].

Climatic variables: Land surface temperature (LST) was extracted from the thermal infrared bands of Landsat imagery and included as an input variable in the modeling framework to account for surface energy dynamics influencing soil salinity. In arid and semi-arid environments, LST serves as a proxy for evapotranspiration and surface moisture conditions, which are intrinsically linked to salt accumulation and redistribution processes [38]. Elevated surface temperatures often correspond to areas with low vegetation cover and high evaporation rates, thereby intensifying salinization through increased capillary rise and subsequent salt deposition on the soil surface [24]. When combined with vegetation indices and spectral salinity indices, LST improves the sensitivity of the model to environmental stress gradients and enables a more nuanced assessment of salinity-prone areas [39]. The inclusion of LST, therefore, complements optical and topographic features by introducing a thermal dimension that enhances model robustness and spatial accuracy.

To enhance the representation of seasonal thermal dynamics in soil salinity modeling, we also incorporated the cumulative LST for both summer (June–August) and winter (December–February) as independent input variables. Because daily LST data at high spatial resolution are not directly available from Landsat due to its 16-day revisit cycle, we implemented MODIS LST product (daily, 1 km resolution) resampling to match Landsat spatial resolution (30 m) using bilinear interpolation and spatial averaging techniques, following the method of Zhang et al. [24]. The obtained dataset enabled the construction of a daily LST time series at a higher spatial resolution. Then, we calculated the seasonal cumulative LST for each year, separately for summer and winter, and integrated these two metrics into the machine learning framework. This approach accounts for cumulative thermal exposure and seasonal temperature extremes, which play a significant role in influencing evaporation rates, capillary rise, and salt crystallization dynamics in the root zone [38].

2.4. Methods

The present study leverages Earth observation data from Landsat 5, 7, and 8 (2000–2025) in combination with ML algorithms to model and monitor soil salinity patterns across Sehb El Masjoune area. Currently, no comprehensive spatial–temporal assessment of soil salinity exists for the area using RS and field data integration. Three common predictive models, i.e., Random Forest (RF), Support Vector Regression (SVR), and Gradient-Boosted Trees (GBT), were trained on field-collected EC data from summer 2022, a period that shows minimal vegetation interference and a high sample data amount (131 points). A total of 44 predictor variables were obtained from RS data, including spectral reflectance bands, salinity indices, vegetation indices, climatic indicators, especially LST with seasonal LST aggregates, and topographic indicators. Following the feature importance analysis, the most influential covariates were retained and incorporated into the spatio-temporal modeling of soil salinization. To ensure a reliable result, all models were trained using the 2022 field ECe dataset, which presents the largest ground truth set, with 65% of the samples used for training, 25% for validation, and 10% testing. Additional independent datasets collected in the summer 2023 (23 samples), 2024 (43 samples), and 2025 (36 samples) were used for testing (Figure 1). The best-performing model was then applied retrospectively to the Landsat archive (2000–2025), limited to summer scenes to ensure temporal consistency. The predicted salinity maps were then discretized into classes and analyzed using an entropy-based Markov transition model, conditioned by hydrological regimes (wet, normal, dry). Finally, the derived regime-specific transition matrices have been used in a pixel-wise probabilistic projection, allowing the soil salinity spatial simulation and associated risk indicators.

2.4.1. Image Pre-Processing

All satellite image preprocessing was conducted using the Google Earth Engine (GEE) platform (https://earthengine.google.com) to make sure that processing across large temporal datasets was consistent and efficient. The analysis focused specifically on the summer imagery for each year, corresponding to the peak of the dry season when salinity-related spectral signals are most pronounced and vegetation cover is minimal, conditions that enhanced the detection of surface salinity [3,40]. Landsat Level-2 Surface Reflectance (SR) products were used whenever available. For earlier periods when SR was unavailable, radiometric calibration was performed by converting digital numbers to top-of-atmosphere reflectance, followed by atmospheric correction using sensor-specific coefficients [41]. Cloud masking was applied using quality assurance bands (QA_PIXEL) to exclude pixels affected by clouds, shadows, or sensor anomalies. Only cloud-free scenes captured in summer with acceptable viewing geometry and radiometric quality were retained. This seasonal standardization helped ensure inter-annual consistency and reduced phenological variability unrelated to salinity dynamics. The resulting summer-based composites were used as input for spectral index calculations and subsequent modeling steps, offering a consistent temporal baseline for long-term soil salinity monitoring.

2.4.2. Indices Calculation

In this study, a comprehensive set of spectral, thermal, and topographic variables was extracted from Landsat imagery to support the spatial modeling of soil salinity (

Table 2. Spectral indices applied in this study for vegetation, soil, and salinity characterization, with their respective abbreviations, full definitions, computational formulas, and bibliographic references.

Index	Full Name	Formula	Reference
NDVI	Normalized Difference Vegetation Index	(NIR − Red)/(NIR + Red)	[43]
DVI	Difference Vegetation Index	NIR − Red	[44]
ERVI	Enhanced Ratio Vegetation Index	(NIR + SWIR2)/2	[45]
ENDVI	Enhanced Normalized Difference Vegetation Index	(NIR + Green – 2 × Blue)/(NIR + Green + 2 × Blue)	[34]
CRSI	Chlorophyll Ratio Stress Index	sqrt((NIR × Red – Green × Blue)/(NIR × Red + Green × Blue))	[46]
MSAVI	Modified Soil Adjusted Vegetation Index	(2 × NIR + 1 − sqrt((2 × NIR + 1)2− 8 × (NIR − Red)))/2	[3]
SAVI	Soil Adjusted Vegetation Index	1.5 × (NIR − Red)/(NIR + Red + 0.5)	[47]
EVI	Enhanced Vegetation Index	2.5 × (NIR − Red)/(NIR + 6 × Red − 7.5 × Blue + 1)	[48]
RVI	Ratio Vegetation Index (Simple Ratio)	NIR/Red	[49]
NDSI	Normalized Difference Soil Index	(Red − NIR)/(Red + NIR)	[3]
SI-T	Soil Index (Temperature-like)	(Red/NIR) × 100	[3]
Albedo	Surface Broadband Albedo	0.356 × Blue + 0.13 × Red + 0.373 × NIR + 0.085 × SWIR1 + 0.072 × SWIR2 − 0.0018	[50]
Albedo_MSAVI	Albedo and MSAVI Composite	sqrt((1 − Albedo)2 + MSAVI2)	[3]
SDI	Salinity Detection Index	sqrt(SI2 + (NDVI − 1)2)	[3]
BI	Brightness Index	sqrt(Red2+ NIR2)	[51]
SI1	Spectral Index 1	sqrt(Blue × Red)	[3]
SI2	Spectral Index 2	(Blue × Red)/Green
SI3	Spectral Index 3	sqrt(Green2 + Red2)
SI4	Spectral Index 4	Blue/Red
SI5	Spectral Index 5	sqrt(Green × Red)
SI6	Spectral Index 6	sqrt(Green2 + Red2 + NIR2)
SI7	Soil/Water Moisture Index	(SWIR1 − NIR)/(SWIR1 − SWIR2)
SI8	SWIR Difference Index	SWIR1 − SWIR2
SI9	SWIR Combination Index	(SWIR1*SWIR2 − SWIR22)/SWIR1
S1	Spectral Ratio 1	Red/Green	[52]
S2	Spectral Ratio 2	Red/Blue	[53]
S3	Spectral Ratio 3	NIR/Green	[44]
S4	Spectral Ratio 4	SWIR1/Green	[45]
S5	Spectral Ratio 5	SWIR1/Red	[54]
S6	Spectral Ratio 6	SWIR1/NIR	[55]
S7	Spectral Ratio 7	SWIR1/Blue	[56]

Note: B, G, R, NIR, SWIR1, and SWIR2 refer to the reflectance of blue band, green band, red band, near-infrared band, and short infrared bands, respectively.

). The modeling approach was based on a supervised learning framework calibrated using data from summer 2022, with corresponding field-measured ECe values. This period was chosen to minimize vegetative interference and maximize the spectral response of salt-affected soils, as recommended in previous remote sensing salinity studies [1,38]. A range of spectral indices, including NDVI, MSAVI, SAVI, NDSI, SI, CRSI, and Albedo (Table 2), were computed according to literature research to capture vegetation stress and soil surface reflectance characteristics, while land surface temperature (LST) and topographic features (elevation, slope, aspect) were added to account for hydrological and thermal controls on salinization [24,36] (Table 2). Once the model was trained–validated on the 2022 dataset and tested on the 2022, 2023, 2024 and 2025 datasets using ground-truth observations, it was retrospectively applied to Landsat imagery from 2000 to 2025, restricted to the summer acquisitions to maintain phenological consistency and avoid seasonal bias. This temporal application strategy allowed the generation of an interannual salinity time series under standardized conditions, supporting a robust analysis of long-term soil degradation trends driven by climatic and anthropogenic factors [39,42].

2.4.3. Machine Learning Algorithms

To model and map soil salinity based on multispectral remote sensing indices, topographic features, and thermal variables, three supervised machine learning algorithms were implemented: RF, GBT and, SVR. These algorithms were selected and chosen because of their capacity to handle nonlinear relationships, manage high-dimensional data, and support flexible feature integration, which are essential characteristics when modeling complex environmental processes such as salinization.

The modeling workflow began with the preparation of input features, including vegetation and salinity indices derived from Landsat reflectance data, LST, and DEM-based parameters (elevation, slope, aspect). All features were calculated for summer 2022 (Figure 2), ensuring phenological consistency with field sampling. Prior to modeling, the variables were standardized where required (particularly for SVR) and examined for multicollinearity. The dataset was then randomly divided into training (65%), validation (25%), and testing (10%) subsets to ensure reliable model evaluation and prevent overfitting.

The Random Forest algorithm [57] was implemented using bootstrap aggregation (bagging), with trees trained on random subsets of the data and a random subset of features at each node split. The RF model was configured to estimate variable importance, and the Gini coefficient was used to assess each feature’s contribution to reducing node impurity during decision tree splits [58]. As highlighted in similar soil salinity studies, the RF model has the robustness to handle noise and feature correlation [59].

The Gradient-Boosting Trees algorithm was applied to construct an ensemble of shallow decision trees in a sequential manner, with each new tree trained to correct the residuals of the previous ensemble [60]. The training procedure relied on gradient descent optimization to minimize the squared-error loss.

Support Vector Regression (SVR) [61] was employed with a Radial Basis Function (RBF) kernel to characterize the nonlinear associations between the RS-derived variables and ECe values. The algorithm was able to implicitly project the input features into a higher-dimensional space thanks to this kernel-based technique, which rendered the intricate relationships linearly tractable for the use of linear regression in that space.

For the Random Forest (RF) model, key hyperparameters including the number of features considered at each split (mtry), the minimum number of terminal nodes, and the number of trees were optimized using a grid search approach. For Gradient-Boosting Trees (GBT), critical hyperparameters such as the learning rate, number of boosting iterations, and tree depth were tuned following a grid-based optimization strategy, consistent with modern gradient boosting implementations [62]. For Support Vector Regression (SVR), the kernel coefficient (γ) and regularization parameter (C) were similarly optimized. All SVR, GBT, and RF models were implemented in Python (version 3.12.12). The final hyperparameter configurations were as follows: the Random Forest model was trained using an mtry value of 10, a total of 500 trees, and a minimum terminal node size of 5. The Gradient-Boosting Tree model was optimized with 500 boosting iterations, an interaction depth of 3, a learning rate of 0.01, and a minimum terminal node size of 5. The SVR model employing an RBF kernel was calibrated using a kernel coefficient γ = 0.1 and a regularization parameter C = 1.

2.4.4. Performance Evaluation

The soil salinity models performance was assessed using four statistical metrics: the coefficient of determination (R²), the root mean square error (RMSE), the mean absolute error (MAE), and the mean bias (Equations (1), (2), (3), and (4), respectively). The R² measures the proportion of variance in the observed salinity explained by the model; values close to 1 indicate strong predictive capacity and a reliable representation of spatial salinity patterns based on input variables such as spectral indices (e.g., SI1, NDVI, MSAVI…), LST, and topography. The RMSE measures the average magnitude of the prediction error, with larger errors penalized more heavily due to squaring, making it useful for identifying overfitting or underfitting in regression models. In contrast, the MAE provides a more robust measure by averaging absolute errors without emphasizing outliers, offering a clearer view of the model’s typical deviation from observed values. Finally, the bias evaluates systematic error by quantifying whether the model consistently over or underestimates soil salinity; values near zero suggest an unbiased prediction. Together, these complementary metrics provide a comprehensive evaluation of model accuracy, generalization, and error structure in environmental prediction tasks [63,64,65,66].

The metrics are defined as follows:

$${\mathbf{R}}^{\mathbf{2}}=\mathbf{1}-{\displaystyle \frac{{\sum }_{\mathrm{i}}{\left({\mathrm{ŷ}}_{\mathrm{i}}-{\mathbf{y}}_{\mathrm{i}}\right)}^{\mathbf{2}}}{{\sum }_{\mathrm{i}}{\left({\mathrm{ȳ}}_{\mathrm{i}}-{\mathbf{y}}_{\mathrm{i}}\right)}^{\mathbf{2}}}}$$

(1)

$$\mathbf{R}\mathbf{M}\mathbf{S}\mathbf{E}=\sqrt{{\displaystyle \frac{\mathbf{1}}{\mathbf{n}}}{\sum }_{\mathbf{i}=\mathbf{1}}^{\mathbf{n}}{\left({\mathbf{y}}_{\mathrm{i}}-{\mathrm{ŷ}}_{\mathrm{i}}\right)}^{\mathbf{2}}}$$

(2)

$$\mathbf{M}\mathbf{A}\mathbf{E}={\displaystyle \frac{\mathbf{1}}{\mathbf{n}}}\times {\sum }_{\mathbf{i}=\mathbf{1}}^{\mathbf{n}}\left|({\mathbf{y}}_{\mathrm{i}}-{\mathrm{ŷ}}_{\mathrm{i}})\right|$$

(3)

$$\mathbf{B}\mathbf{i}\mathbf{a}\mathbf{s}={\displaystyle \frac{\mathbf{1}}{\mathbf{n}}}\times {\sum }_{\mathbf{i}=\mathbf{1}}^{\mathbf{n}}({\mathrm{ŷ}}_{\mathrm{i}}-{\mathbf{y}}_{\mathrm{i}})$$

(4)

where ${\mathbf{y}}_{\mathrm{i}}$ is the measured soil salinity value, ${\mathrm{ŷ}}_{\mathrm{i}}$ is the soil salinity inversion value, and n is the number of sample points.

2.4.5. Spatial Uncertainty Quantification of SVR-Based Soil Salinity Predictions Using Standard Deviation (SD) Across Ensemble Iterations

To quantify the spatial uncertainty associated with the Support Vector Regression (SVR) model predictions, a standard deviation (SD)-based approach was adopted. This method evaluates the variability of predicted ECe at each pixel by analyzing the dispersion across 50 independent prediction iterations. For each iteration i, a complete prediction map $P_{\mathrm{i}}\left(x, y\right)$ was generated using resampled training subsets. The uncertainty at each pixel location (x, y) was then computed as the standard deviation of predicted values across all iterations:

$$\mathrm{S}\mathrm{D}\left(\mathrm{x},\mathrm{y}\right)=\sqrt{\left(1/\left(\mathrm{N}-1\right){\sum }_{\mathrm{i}=1}^{\mathrm{N}}{\left({\mathrm{P}}_{\mathrm{i}}\left(\mathrm{x},\mathrm{y}\right)-\mathrm{P}\overline{\left(\mathrm{x},\mathrm{y}\right)}\right)}^2\right)}$$

(5)

where N = 50 is the number of iterations, and $P\overline{\left(x,y\right)}$ is the mean predicted value. The resulting SD map reflects the prediction variability at each location, with lower SD indicating stable, confident model predictions and higher SD revealing areas of greater uncertainty. This ensemble-based approach, in accordance with de Mello et al. [67] and Helfenstein et al. [68], provides a robust and interpretable measure of prediction reliability without assuming specific error distributions.

2.4.6. Transition Estimation from Class Shares and Hydrologic Conditioning

Soil salinity was discretized into six ECe classes: 0–4, 4–8, 8–16, 16–32, 32–64, and ≥64 dS/m. Let $r(t)= {\{r_i(t)\}}_{i=1}^6$ and $r(t+1)= {\left\{r_j\right(t+1\left)\right\}}_{j=1}^6$ denote the vectors of proportional area (i.e., relative frequency of pixels) for the six salinity classes in years $t$ and $t+1$, respectively, such that ${\sum }_{i=1}^6{r}_i(t)={\sum }_{j=1}^6{r}_j(t+1)=1$. Year-to-year flows $T_{ij}\left(t\right)$ (fraction of area moving from class i at t to class j at t + 1) were reconstructed by solving a cross-entropy problem with a diagonal-biased prior:

$$mi{n}_{T\ge 0}{\Sigma }_{i,j}{T}_{ij}\left(t\right)\times log{\displaystyle \frac{T_{ij}\left(t\right)}{Q_{ij}}}, s.t.{\Sigma }_j{T}_{ij}\left(t\right)=r_i\left(t\right),{ \Sigma }_i{T}_{ij}\left(t\right)=r_j\left(t+1\right)$$

(6)

with ${\mathcal{Q}}_{ij}\propto e^{-\alpha |i-j|}$ and $\alpha =0.8$ to favor adjacent-class changes while allowing larger moves. The optimizer used the Iterative Proportional Fitting Procedure (IPFP), also known as the Sinkhorn algorithm, alternately scaling rows and columns until the marginal constraints matched r(t) and r(t + 1) [69,70].

A row-stochastic instantaneous transition matrix was then obtained:

$$P_{ij}\left(t\right) = {\displaystyle \frac{T_{ij}\left(t\right)}{{\sum }_k{T}_{ik}\left(t\right)}}.$$

(7)

Hydrologic regimes were assigned from SPI-12: WET for SPI ≥ +0.5, DRY for SPI ≤ 0.5, and NRM otherwise. Each transition $t \to t+1$ inherited the regime of year t. Regime-specific matrices were formed by aggregating flows within each regime and row-normalizing, consistent with entropy-based transition modeling frameworks [71].

$$P_{ij}\left(Y\right)={\displaystyle \frac{{\sum }_{t\in Y}{T}_{ik}\left(t\right)}{{\sum }_{t\in Y}{\sum }_k{T}_{ik}\left(t\right)}}\left(Y\in \{\mathrm{WET},\mathrm{NRM},\mathrm{DRY}\}\right),$$

(8)

and approximate standard errors followed a multinomial form:

$$\mathrm{SE}\left(\widehat{p_{ij}}\right) \approx \sqrt{{\displaystyle \frac{\widehat{p_{ij}}\left(1-\widehat{p_{ij}}\right)}{n_i}}}, \widehat{p_{ij}}=P_{ij}\left(Y\right), { n}_i={\sum }_{t\in Y}{\sum }_k{T}_{ik}\left(t\right)$$

(9)

These regime-conditioned P(Y) matrices were supplied to the mapping stage.

2.4.7. Pixel-Wise Probabilistic Projection and Spatial Mapping

The regime-specific transition matrices P(Y) derived from class-based dynamics were subsequently applied at the pixel level to simulate the spatial evolution of soil salinity. The baseline year (2018) was selected as the initial state because it represented the most recent and hydrologically well-documented wet year, during which extensive leaching processes resulted in a balanced and representative distribution of soil salinity across the study area. This year therefore provided an optimal reference condition, with a fair proportion of soils exhibiting ECe values within the suitable range for agricultural production. The Markov chain framework was then used to project class transitions over a five-year horizon (H = 5), following the observed hydrological sequence of wet, normal, and dry years between 2019 and 2025.

For each pixel, the class probability vector at step k was calculated as

$${\pi }^{\left(k\right)} = {\mathrm{e}}_c {\mathrm{P}}_{t_0+1} {\mathrm{P}}_{t_0+2} \cdots {\mathrm{P}}_{t_0+k},$$

(10)

where ${\mathrm{e}}_c$ represents the one-hot vector corresponding to the initial class, and each $P_t$ corresponds to the transition matrix associated with the hydrological regime of year t. This formulation allowed the estimation of transition probabilities for every pixel under a realistic sequence of climatic regimes, rather than a stationary or uniform assumption.

From the probabilistic trajectories, three key indicators were derived to capture soil salinity dynamics. The risk of high salinity (≥32 dS/m) quantified the average likelihood of a pixel remaining or shifting into the upper salinity classes, indicating degradation risk. The improvement potential (≤8 dS/m) represented the highest probability of recovery toward low-salinity states, while the extreme salinity probability (≥64 dS/m) denoted the chance of reaching irreversible salinization. These indicators jointly describe persistence, reversibility, and transition likelihood under variable hydrological regimes.

To support sustainable land management, all identified opportunity classes were treated collectively as priority zones requiring protection and proactive intervention to mitigate ongoing or potential salinization. This integrated framework emphasizes preserving soil productivity through targeted actions that prevent further degradation and promote long-term resilience.

Finally, a drought-stress simulation was conducted to estimate the cumulative probability of exceeding critical salinity thresholds (≥16, ≥32, or ≥64 dS/m) under prolonged dry conditions. This was achieved by repeatedly applying the dry-year transition matrix over ten consecutive iterations (K = 10) and combining annual exceedance probabilities using a union approximation:

$$P\left(\mathrm{any} \ge X \mathrm{within} K\right)\approx 1-{\prod }_{k=1}^K\left(1-P_k\left(\ge X\right)\right)$$

(11)

This stress test identifies areas that are likely to become irreversibly saline if multi-year droughts persist, thereby delineating zones where preventive measures are most urgent. Overall, the combination of transition-based projection and probabilistic mapping provides a dynamic and spatially explicit framework for diagnosing soil salinity trajectories, assessing vulnerability to hydrological stress, and prioritizing management interventions under changing climatic regimes [71].

3. Results

3.1. Soil Salinity Predictors

A comprehensive correlation analysis was conducted to explore the relationships between ECe and various environmental variables, including spectral reflectance bands, salinity indices, vegetation indices, topographic attributes, climatic variables, and composite indicators (Figure 3). As shown in similar salinity modeling studies [3], EC had a low direct correlation with individual spectral bands (Panel B), indicating limited sensitivity of raw surface reflectance to salinity variations. However, climatic variables, particularly summer land surface temperature (LST_Summer), showed moderate positive correlations with EC (r = 0.53), suggesting a potential link between surface heating and evaporation in salt concentration during the dry season.

Among the topographic factors, elevation showed a significant negative correlation (r = −0.56), indicating that salinity tends to accumulate in lower-lying areas due to runoff and capillary rise, consistent with previous terrain salinity relationships observed in arid regions [72]. Correlations with slope (r = 0.10) and aspect (r = −0.16) were weaker but supported the topographic influence on water and salt redistribution.

Salinity-specific indices demonstrated moderate correlations with EC. Notably, SI.T, NDSI, SI8, and SI9 were among the most correlated, confirming the effectiveness of these indices in capturing the spectral signatures of saline surfaces. Similarly, vegetation indices such as CRSI (r = −0.47) and ENDVI (r = −0.40) were negatively correlated with EC, which is consistent with the ecological response where higher soil salinity reduces vegetation vigor.

Overall, the correlation matrix informed the feature selection process, whereby highly redundant variables were eliminated, and only the most relevant predictors from each category (e.g., CRSI, ENDVI, NDSI, SI8, LST_Summer, and elevation) were retained for the model input. This procedure ensured multicollinearity reduction and improved model generalization, in accordance with methodologies employed in similar remote sensing-based salinity assessment.

3.2. Model Accuracy Validation and Interpretation

The predictive performance of GBT, RF, and SVR was assessed using standard accuracy metrics, including the R², RMSE, MAE, and bias. As shown in

Table 3. Performance comparison of machine learning models (GBT, RF, and SVR) for soil salinity prediction based on training and testing datasets, showing the coefficient of determination (R²), root mean square error (RMSE, dS/m), mean absolute error (MAE, dS/m), and bias.

Model	Train				Test
	R2	RMSE (dS/m)	MAE (dS/m)	BIAS	R2	RMSE (dS/m)	MAE (dS/m)	BIAS
GBT	0.89	22.66	14.93	0.43	0.65	40.07	28.23	14.55
RF	0.95	15.95	10.78	0.74	0.64	40.56	29.40	12.32
SVR	0.90	22.27	13.93	−0.58	0.76	32.91	23.12	8.47

and Figure 4, the SVR model exhibited superior performance with an R² of 0.90 and 0.76 for train and test, respectively, indicating its strong predictive capability and minimal deviation from the observed values. The SVR model, known for its robustness in handling high-dimensional datasets, demonstrated strong predictive performance, with training metrics of R² = 0.90, RMSE = 22.27 (dS/m), and MAE = 13.93 (dS/m). The model exhibited a slight negative bias to a low positive (−0.58 and 8.47 for train and test, respectively), suggesting a minor underestimation tendency in salinity prediction. Its performance remained consistent across the testing phase, yielding the highest accuracy among the evaluated models (including GBT and RF), highlighting its superior generalization capability and reliability for soil salinity estimation. The GBT and RF models also performed well, with coefficients of determination (R²) of 0.65 and 0.64, respectively. However, they exhibited slightly higher error values, with RMSEs of 40.07 and 40.56 (dS/m) and MAEs of 28.23 and 29.40 (dS/m), respectively, indicating a minor tendency toward overestimation. Despite the discernible divergence between the calibration and validation phases, the degree of variability remains well within the acceptable threshold (around 0.70 R²), indicative of stable modeling behavior. The retention of substantial explanatory capability on independent sets underscores the models’ generalization effectivity, even when trained on a relatively constrained calibration set like the 2022 set. This stability suggests that the used labeled data sufficiently encapsulated the underlying salinity–covariate dynamics, enabling robust predictive inference.

3.3. Feature Importance Analysis

A feature relative importance graph is shown in Figure 5. Notably, feature importance analysis based on the three used models indicated that LST_summer, elevation, S3, CRSI, and RVI were among the most significant predictors, highlighting the combined effects of topography, surface heating, and vegetation degradation. The negative correlation with elevation (r = −0.56) and vegetation indices such as ENDVI (r = −0.40) suggests that salinization is more severe in lower-elevation zones and sparsely vegetated fields, confirming hydrological concentration zones and limited buffering from the canopy cover.

3.4. Spatial and Temporal Variability (2000–2025)

Using the SVR model, validated as the most accurate (R² = 0.76), soil salinity maps generated from 2000 to 2025 revealed an overall trend of salinization intensification across the landscape (Figure 6). The multi-date classification maps show that areas with ECe ≥ 32 dS/m, as well as the most saline class (≥64 dS/m), are systematically concentrated along low-lying depressions and former lake boundaries and undergo repeated expansions during dry phases with only partial retraction in wetter years. This behavior is quantified by a time series of the total area with ECe ≥ 32 dS/m (Figure 7), which fluctuates between roughly 7000 and 12,500 ha over the study period, with a pronounced minimum around 2008 and several episodes of expansion in the early 2000s, 2011–2012, and 2019–2022. In contrast, the strongly and extremely saline classes ([32–64] and [≥64] dS/m) display pronounced interannual variability but only weakly negative trends, indicating that the net salinization signal over 2000–2025 is dominated by a broad shift from low-to-moderate salinity, while the most extreme salinity levels remain largely confined to persistent hotspots rather than undergoing sustained long-term expansion.

The local hydrological processes contributing to salinization were clarified by a field water sampling campaign carried out in March 2025 as a supplement to the remote sensing analysis. The impact of shallow saline water was validated by preliminary in situ EC measurements and water table observations, especially in regions where salinity intensification occurred. According to Demelash et al. [37], post-irrigation soil profiles frequently exhibit elevated EC levels, particularly when using conventional irrigation regimes with constrained drainage capacity. Recent research in Central Asia and Eastern Africa has recommended adaptive management strategies that combine vegetation restoration, targeted drainage, and spatially explicit monitoring to address this issue [3,37].

3.5. Spatial Distribution of Predicted Soil Salinity Classes and Associated Uncertainty Derived from SVR-Derived Modeling

Panel A in Figure 8 displays the spatial distribution of soil salinity classes (0–4 to ≥64 dS/m) across the study area, as predicted by the Support Vector Regression (SVR) model. The map reveals a pronounced salinity gradient extending from the central sebkha depression, characterized by extremely saline soils (≥64 dS/m, shown in deep red), to the surrounding agricultural zones with lower salinity levels (0–8 dS/m, shown in blue tones). The delineated salinity classes highlight the strong spatial organization of salinization processes, controlled by drainage conditions, micro-relief, and groundwater discharge.

Panel B illustrates the spatial variability of prediction uncertainty, expressed as the standard deviation (SD) derived from 50 independent SVR iterations. The SD values ranged from approximately 2.6 to 26.7 dS/m, exhibiting clear spatial gradients across the study area. The highest-uncertainty zones (red tones) are concentrated along the central saline depression and near the class boundaries delineated in Panel A, reflecting unstable model predictions in heterogeneous or transitional soil environments. These areas likely correspond to regions of high surface reflectance variability, mixed land-cover composition, and irregular salt crust formation, where the spectral response is less consistent across scenes. Conversely, lower SD values (blue to green tones) prevail in well-drained agricultural fields and vegetated zones, indicating stable spectral characteristics and stronger model generalization. The spatial pattern of uncertainty further emphasizes the influence of topography and hydrological setting: elevated SD values along the sabkha margins suggest that minor elevation differences and micro-relief variations enhance local salinity heterogeneity. Overall, the SD map provides a spatially explicit representation of model confidence, supporting both the identification of zones requiring targeted ground verification and the interpretation of salinity dynamics under complex hydro-sedimentary conditions.

3.6. Temporal Assessment of Salinization as a Function of Drought

The temporal evolution of soil salinity classes between 2000 and 2025 reveals distinct but gradual shifts across the mapped EC ranges under varying hydroclimatic regimes. Based on the 12-month Standardized Precipitation Index (SPI-12), years with SPI ≥ +0.5 were classified as wet, SPI ≤ −0.5 as dry, and intermediate values as normal, following the World Meteorological Organization’s (WMO) guidelines for consistent meteorological drought and wetness classification. Precipitation terciles served as secondary validation criteria, while SPI remained the principal indicator when |SPI-12| ≥ 1 (Figure 9).

Hydroclimatic variability plays a central role in shaping these dynamics (Figure 10). Dry years (SPI ≤ −0.5) are associated with elevated proportions of moderately to strongly saline soils (8–32 dS/m), due to increased evapotranspiration, reduced leaching, and upward capillary rise in salts. In contrast, wet years (SPI ≥ +0.5) temporarily reduce surface salinity, particularly within the 0–8 dS/m range, through enhanced infiltration and dilution processes that promote salt leaching below the root zone. However, this relief is transient, as subsequent dry cycles reaccumulate salts through evaporation and poor drainage. Lagged-climate regressions further support these findings: SPI-1_t−1 is negatively associated with the 4–8 dS/m class (p < 0.05), indicating that short wet pulses enhance leaching of lightly saline soils; SPI-6_t−1 correlates positively with the 32–64 dS/m class (p < 0.05), suggesting that seasonally wet antecedents can raise groundwater tables and trigger salinity rebound; and SPI-12_t−1 is negatively related to 16–32 dS/m (p ≤ 0.05), confirming that extended wet conditions suppress mid-band salinity in the following year.

Over the 2000–2025 period, the linear regression-derived trends in the annual area shares of the six ECe classes reveal a clear redistribution pattern. The non-saline (0–4 dS/m) and slightly saline (4–8 dS/m) classes exhibit moderate declines of −1.17 and −0.22 percentage points per decade (pp.dec⁻¹), respectively, indicating a contraction of the low-salinity range (0–8 dS/m, −1.39 pp.dec⁻¹). This reduction likely reflects progressive salt accumulation and reduced leaching efficiency in response to intensified evapotranspiration and irrigation recycling. Conversely, the moderately saline (8–16 dS/m) and strongly saline (16–32 dS/m) classes increased by +0.73 and +1.58 pp.dec⁻¹, respectively, highlighting an expansion of moderately affected areas driven by salt mobilization during dry hydrological years. Higher salinity classes (32–64 dS/m and ≥64 dS/m) declined slightly (−0.16 and −0.76 pp.dec⁻¹ respectively), suggesting localized persistence of severe saline patches without significant spatial spread (Figure 10).

Overall, the data reveals a slow but concerning redistribution from non-saline to moderate salinity conditions, under a cyclical but progressively intensifying salinization trend. Recurrent drought periods remain the dominant driver sustaining and amplifying soil salt concentrations over time, confirming both the robustness of the adopted methodology and the reliability of the selected climatic covariates in modeling soil salinity dynamics across the region.

3.7. Spatial Probability of Salinization Risk

The regime-specific transition matrices (Figure 11) show clear differences in salinity persistence and inter-class exchange under normal, dry, and wet conditions. Across all regimes, diagonal elements dominate, indicating substantial year-to-year persistence within salinity classes. Persistence probabilities generally range from approximately 29–54% for low and moderate classes (≤16 dS/m) and increase to about 42–58% for higher salinity classes (16–32, 32–64, and ≥64 dS/m).

Off-diagonal transitions are primarily concentrated between adjacent salinity classes (e.g., 4–8 ↔ 8–16 and 8–16 ↔ 16–32 dS/m). Transitions spanning more than one class interval are consistently low across all regimes. For moderate salinity levels, upward and downward transitions exhibit similar magnitudes under all hydrological conditions. For example, the probability of a transition from 16–32 to 8–16 dS/m is 19% under normal conditions, 22% under dry conditions, and 20% under wet conditions.

For higher salinity classes (≥16 dS/m), regime-dependent differences are evident. In dry years, pixels in the 16–32 and 32–64 dS/m classes show slightly higher probabilities of downward transitions than upward transitions, while maintaining strong class persistence. In contrast, under wet conditions, the transition matrices show a shift in exchange dynamics relative to normal years. Two patterns are observed. First, for salinity classes below 32 dS/m, downward transition probabilities tend to exceed upward ones. For instance, the probability of moving from 4–8 dS/m to 0–4 dS/m is 24%, whereas the reverse transition is 15%. Similarly, the 8–16 to 4–8 dS/m transition increases by 11 percentage points under wet conditions compared to normal. This pattern suggests a tendency toward net freshening in the low-to-moderate salinity range during wet periods. Second, wet years are associated with a decrease in the probability of transitions from low-to-moderate salinity classes into the highest class, ≥64 dS/m. For instance, transitions from 8 to 16 dS/m decrease by two percentage points and from 16 to 32 dS/m by one percentage points relative to normal years. Overall, these results indicate that wetter conditions modestly favor salinity reductions at lower salinity levels and slightly limit shifts into the most saline state. Despite these differences, the highest salinity class (≥64 dS/m) remains highly persistent across all regimes, with diagonal probabilities exceeding 50%.

Spatial probability maps reveal distinct patterns of salinity dynamics and susceptibility across the study area (Figure 12). The mean probability of exceeding 32 dS/m (panel A) highlights the zones most prone to moderate-to-severe salinity accumulation, with the central depression exhibiting the highest likelihoods. The probability of reaching extreme salinity (≥64 dS/m; panel B) follows a similar spatial trend but with more confined hotspots, largely corresponding to poorly drained lowlands. Conversely, the probability of improvement to ≤8 dS/m (panel C) is greater along the northern and western peripheries, indicating areas with stronger recovery potential. Panels D–F show the cumulative probability of exceeding 16 dS/m under two, three, and five consecutive dry years, respectively. The maps depict a progressive expansion of high-risk zones with the duration of drought sequences, especially toward the central and downstream sectors.

4. Discussion

The Sehb El Masjoune basin, located in the semi-arid region of central Morocco, exhibits biophysical characteristics that are highly conducive to salinization processes. These characteristics include a closed drainage system, shallow aquifers, and high evaporative demand. Such conditions are similar to those observed in other saline inland depressions, such as the Weiku Oasis in Xinjiang [18] and the northeastern Nile Delta [4], where salinity results from inadequate drainage, saline groundwater upsurge, and inefficient irrigation practices. In Sehb El Masjoune, the interplay of climatic aridity, topographic depressions, and historical irrigation practices has created hydrological concentration zones where soluble salts accumulate and persist. This topographically driven mechanism of salt redistribution is consistent with observations in the Lake Urmia basin [59] and the arid basins of southern Xinjiang [3], where lower-elevation areas act as salt sinks, thereby intensifying secondary salinization.

4.1. Models Potentially and Contextualization

The study assessed three machine learning algorithms, SVR, RF, and GBT, for soil salinity prediction. This allowed the selection of the optimal model for both spatial and temporal modeling of salinization in the studied semi-arid abandoned agricultural lands. Accordingly, SVR achieved the best predictive accuracy (R² = 0.76), followed by GBT (R² = 0.65), using the test set (Table 3). These results align with similar ML evaluations in Xinjiang and Iran [3,21], where SVR consistently outperformed ensemble tree methods for nonlinear soil and salinity datasets. Zhao et al. [21] reported an SVM R² = 0.756 in southern Xinjiang, higher than RF and GBT, while Chen et al. [73] demonstrated that SVM-based ensemble learning captured the spatial heterogeneity of soil salinity more effectively than tree-based classifiers. Comparable findings were also noted by Mirzaee et al. [59] for Lake Urmia and by Zhang et al. [3] for southern Xinjiang, confirming that kernel-based algorithms efficiently model salinity under variable spectral and topographic conditions. The superior accuracy of SVR is attributable to its ability to manage multicollinearity among spectral indices and to generalize over heterogeneous sampling environments.

4.2. Evident Climate Impact on Salinity Dynamics in the Studied Semi-Arid Landscape

Feature importance analysis (Figure 5) identified elevation, LST_summer, S3, CRSI, and RVI as the most significant predictors. This combination includes terrain, thermal, and vegetation stress information, which collectively play a crucial role in salinity mapping [18,74]. The negative correlations between salinity and both elevation (r = −0.56) and vegetation indices (ENDVI, r = −0.40) suggest that salinization is more pronounced in low-lying, sparsely vegetated areas, a pattern also observed in the Yellow River Delta [75] and grassland mining regions of Inner Mongolia [74]. The key role of LST as an indicator of soil salinity aligns with the findings of Zhao et al. [21] and Zhang et al. [3], which emphasized that surface temperature values indicate soil crusting and diminished moisture retention, thereby promoting salt accumulation. Furthermore, the prominence of ENDVI and CRSI supports the effectiveness of combined spectral–vegetation metrics for monitoring soil degradation in arid croplands [73,76]. Topography also influences salt migration and concentration, as evidenced by studies in the Nile Delta [4] and southern Xinjiang [3], where lower elevations were associated with shallow saline groundwater and elevated ECe values. Thus, the identified predictors effectively capture both the direct and indirect biophysical processes governing salinity in Sehb El Masjoune.

The spatiotemporal analysis revealed a systematic escalation in soil salinity levels from 2000 to 2025 (Figure 6), with a notable emphasis on areas characterized by low elevation and inadequate drainage. The temporal framework anchored in summer mitigated phenological variability and ensured seasonal uniformity throughout the 26-year duration, aligning with the methodological strategies advocated by Chen et al. [73] and Guo et al. [75]. Climatic fluctuations, predominantly characterized by deficits in precipitation and rising land surface temperatures, account for a significant portion of the temporal variability observed in salinity concentrations. Parallel correlations between climatic influences and salinity have been documented in the Lake Urmia basin [59] and the Yellow River Delta [75]. In these investigations, salinity maxima were found to coincide with warmer and drier years, thereby illustrating the impact of evaporation-induced capillary rise (Figure 9).

4.3. Further Validation of Used Model and Its Scalability Using Multi-Temporal Datasets

Field validation conducted in 2023, 2024, and 2025 substantiated the modeled spatiotemporal dynamics (Figure 8), with measured electrical conductivity values corroborating the predictive accuracy and temporal robustness of 2022 using the SVR model. This reaffirms its capacity for temporal extrapolation and aligns with findings by Mirzaee et al. and Zhang et al. [3,59], who indicated that machine learning models trained on heterogeneous datasets integrating climatic, topographic, and spectral indices datasets maintain validity when essential predictors (e.g., LST, elevation, ENDVI, NDSI) exhibit interannual variability. The persistence of high-salinity zones within hydrologically isolated depressions further supports the hypothesis that restricted leaching and shallow groundwater dynamics are primary long-term drivers of surface salt accumulation [4,21]. Collectively, these results emphasize the model’s relevance for continuous monitoring frameworks and advocate for the integration of SVR ensemble architectures within operational salinity early-warning systems, as demonstrated in recent studies from Xinjiang [3] and northern Iran [76]. The integration of earth observation datasets and ML offers promising paths for enhancing soil salinity monitoring at both regional and temporal scales. This study demonstrated the relative effectiveness of the SVR model in capturing the interplay between climatic, spectral, topographic, and vegetation-related factors driving salinization. The retrospective modeling from 2000 to 2025, combined with multiyear Landsat data and in situ validation, provides a high-resolution framework for long-term salinity surveillance.

4.4. Implications for the Spatial and Temporal Salinization Dynamics Analysis

The transition probabilities provide quantitative evidence of how climate-driven hydrological variability governs soil salinity dynamics (Figure 11). The higher persistence observed under normal years reflects an equilibrium between capillary rise and limited leaching, while wet years promote vertical solute redistribution and partial recovery of moderately saline soils. This finding is supported by Rengasamy [77], who demonstrated that shallow saline groundwater and evapotranspiration maintain a steady upward salt flux in arid regions. Conversely, the dry-year matrices highlight the fragility of this equilibrium, as even short drought sequences can push marginally saline soils beyond agronomic thresholds, reflecting salt accumulation through evaporation and capillary rise under reduced precipitation and elevated PET [78,79,80,81,82]. The persistence of the ≥64 dS/m class across all regimes further demonstrates the structural resilience of extremely saline soils, controlled by fine-textured sediments, low drainage capacity, and flat micro-relief, conditions similar to those described by Sun et al. [83] in endorheic aquifers of central Mexico.

This asymmetry between recovery and degradation processes highlights the nonlinearity of salinity evolution, where degradation under water stress is faster and more probable than recovery during wet cycles [10]. Such behavior underscores the need for proactive management during early salinization stages, emphasizing drainage maintenance, and timely leaching interventions. The economic implications of delayed reclamation are well illustrated by Qadir et al. [84], who demonstrated that the cost of inaction far exceeds the investment required for early restoration. Moreover, the steeper transition gradients observed under dry regimes confirm the cumulative impact of consecutive water deficits on salinity intensification [3,69], reinforcing the vulnerability of semi-arid systems to climate variability.

These results are consistent with hydro-salinity theory for endorheic basins, where hydrological cycles alternate between leaching and concentration phases. During water-limited periods, shallow saline groundwater and strong evaporative demand drive upward capillary fluxes that deposit salts near the surface [77]. In contrast, wetter conditions enhance percolation and lateral redistribution of salts, though rarely leading to true desalinization due to the lack of external drainage. The study by Daliakopoulos et al. [79] across Mediterranean basins aligns with this behavior, showing that climatic oscillations and limited drainage capacity produce cyclic salinity buildup and only temporary recovery. The matrices therefore offer a probabilistic expression of the physical salt cycle, with wet phases promoting partial salt leaching, while subsequent dry years reconcentrate salts in the root zone [10].

The persistence of high salinity even during favorable years underscores the structural “stickiness” of the system. Hydrogeologic convergence toward central depressions, combined with low permeability and shallow water tables, restricts real salt export [83]. In the observed long-term pattern, a decline in the ≥64 dS/m and 32–64 dS/m fractions with a rise in the 8–32 dS/m class indicates internal redistribution rather than overall desalinization (Figure 10). Similar findings were reported by Wang et al. [78] in the Weigan Oasis, where surface desalinization was offset by deeper salt accumulation, confirming vertical reorganization rather than salt loss. The lagged SPI associations further corroborate this cyclicity: positive SPI-6_t−1 values indicate salinity rebound following short wet intervals, while negative SPI-12_t−1 coefficients correspond to temporary recovery after prolonged wet phases. These results agree with Elair et al. [85], who documented alternating wet and dry periods in the Marrakech–Safi region using SPI-6 and SPI-12, confirming the strong coupling between rainfall deficits, vegetation stress, and hydrological drought across semi-arid basins, which directly impact the soil salinity level [86].

From a management perspective, the 16–32 dS/m range emerges as the most sensitive leverage zone, showing both responsiveness to wet conditions and rapid degradation under dryness. Targeted interventions, such as timed leaching and subsurface drainage, have proven effective in stabilizing this class, as confirmed by Li et al. [71] in Xinjiang. In contrast, low-salinity soils (0–4 dS/m) demand protective practices, such as mulching, salt-tolerant rotations, and seepage control, to prevent salinity encroachment during drought cycles [1]. The overall evidence emphasizes that salinity mitigation in Sehb El Masjoune and similar basins depends less on occasional rainfall than on maintaining continuous hydrological connectivity that facilitates salt export [84]. Finally, the probabilistic patterns captured in this study reflect the core hydro-salinity feedbacks that shape semi-arid landscapes. Elevated probabilities of exceeding 32 and 64 dS/m correspond to geomorphic depressions where capillary rise and evaporation dominate, while marginal zones benefit from episodic leaching through lateral flow. The progressive salinity intensification observed during successive dry years quantifies the system’s sensitivity to aridity and limited drainage. These findings, consistent with Rengasamy and Daliakopoulos et al. [77,79], suggest that sustained management, including efficient drainage, shallow water-table control, and protection of low-salinity fringes, is essential to interrupt cyclic salt accumulation. Although leaching can temporarily reduce surface salinity, its effects remain short-lived unless supported by long-term hydrological control and adaptive irrigation scheduling [87,88].

4.5. Challenges and Future Perspectives

Several challenges and research opportunities remain for future exploration. One of the key prospects is the expansion to finer spatial and temporal resolutions. The current use of Landsat imagery, while adequate for regional-scale modeling, may miss sub-field heterogeneity critical for precision agriculture. The integration of Sentinel-2 MSI, PlanetScope, or hyperspectral UAV data can enhance the sensitivity of salinity indices and enable early detection of saline stress, particularly in mosaicked or fragmented fields. Likewise, coupling temporal fusion techniques to derive daily surface reflectance and thermal metrics could improve the responsiveness of models to seasonal or irrigation-related changes. Second, future work could benefit from multi-depth salinity inversion, particularly when paired with soil moisture and groundwater monitoring. Although this study focused on surface salinity estimation (0–15 cm), the inclusion of soil profile ECe data could help develop vertically stratified models. Such approaches are essential in irrigated zones where leaching, capillary rise, and subsurface salt storage vary with depth and season. A further challenge lies in model generalizability and transferability. While the SVR model performed well in the Sehb El Masjoune context, its applicability to different agroecological zones or years with anomalous climatic conditions (e.g., drought, floods) remains uncertain. Additionally, socio-economic and management data remain underutilized. Incorporating irrigation practices, cropping patterns, and drainage infrastructure into ML models could improve prediction accuracy and support actionable land management strategies. Future frameworks should integrate participatory mapping, ground sensors, and RS to create feedback loops with farmers and water managers. Future work should strive to enhance vertical and temporal resolution, integrate multi-source data, and strengthen links between model outputs and land use policy to enable scalable, sustainable, and site-adapted salinity management.

5. Conclusions

This study provided a reliable implication via a comprehensive assessment of soil salinity dynamics over the last quarter-century in the Sehb El Masjoune area by integrating field observations, RS data, and ML. Among the three ML models, their accuracies are ranked as SVR > GBT > RF, which shows the outperformance of the SVR algorithm in modeling soil salinity at an abandoned agricultural saline soil-characterized area. The summer observation-based approach ensured phenological consistency across years, and the feature importance analysis revealed that elevation, LST_Summer, S3, CRSI, and RVI were the most influential predictors of salinity variation. The SVR model was selected for its superior predictive accuracy, and we developed a robust spatiotemporal framework to map and monitor salinity based on Landsat-derived spectral indices, LST, topographic variables, and composite indicators. The resulting salinity maps and trend analyses highlighted a significant increase in saline zones (16–32 dS/m) over time, particularly in low-lying and poorly drained areas, with transitions from moderate to severe salinity classes confirmed by classification shifts and area distribution curves. Over the past quarter-century, the region’s soils have undergone a quiet transformation, gradually shifting from largely non-saline conditions toward predominantly moderate salinity levels, with localized but persistent hotspots of severe salinity. This creeping change reflects the accumulated impact of climate variability and land and water management, where dry spells actively concentrate salts through enhanced evapotranspiration and reduced leaching, while wet periods provide only temporary relief. The data reveal a concerning long-term trajectory in which the landscape is slowly losing its natural resilience against salinization pressures. Multiple independent diagnostics all point to the same conclusion: probability surfaces localize high risk (≥32 and ≥64 dS/m) within the central depression, expanding under multi-year drought sequences while peripheral zones retain greater recovery potential; temporal diagnostics based on SPI (the Standardized Precipitation Index) show drought intensification after 2016 and a redistribution of area from 0–8 dS/m to 8–32 dS/m; regime-specific transitions indicate relative stability in normal years, modest net improvement in wet years consistent with partial recovery within low-to-moderate salinity classes, and net salinity build-up in dry years arising from the balance of inflows into and persistence of moderate-to-high salinity classes rather than from uniformly dominant upward transitions; and spatial analyses reveal a pronounced gradient from the Sabkha core toward surrounding croplands, with the greatest uncertainty concentrated along class boundaries and Sabkha margins. Collectively, these results indicate a continuous shift toward moderate salinity with isolated extreme hotspots, underscoring the need to prioritize mitigation and monitoring in low-lying, poorly drained sectors and drought-sensitive margins.

By combining long-term earth observation data with ground-truth validation and machine learning algorithms, this study contributes to a scalable and transferable methodology for salinity monitoring in arid and semi-arid agricultural regions. The findings underline the urgency of implementing adaptive soil and water management strategies, particularly in salinity-prone landscapes facing climatic stress, irrigation inefficiencies, and land degradation. The integration of such data-driven tools into regional land planning and irrigation governance could enhance early-warning systems, optimize resource allocation, and support sustainable agricultural development under increasing environmental constraints.