GIS-Based Assessment of Grassland Dynamics and Sustainable Pasture Planning in South and Nabatieh Governorates, Lebanon (2000–2024)

Abstract

Sustainable pasture management is essential for preserving ecosystem services and supporting livestock production in Mediterranean environments experiencing increasing land-use pressure and climate variability. This study evaluates grassland dynamics in South Lebanon between 2000 and 2024 using multi-temporal Landsat imagery integrated with Geographic Information Systems (GISs) and Random Forest classification. Historical Landsat surface reflectance products were processed using standardized preprocessing procedures, cross-sensor harmonization, and supervised classification. Classification accuracy was assessed using independent validation samples, yielding overall accuracy values ranging from 91.8% to 98.6% and Kappa coefficients between 0.918 and 0.975. Pasture areas were subsequently classified into dense and dispersed grasslands using quantitative NDVI thresholds. A pixel-based transition matrix was employed to quantify land-cover persistence, gains, losses, and conversions throughout the study period. The results reveal substantial spatial redistribution of grasslands associated with agricultural intensification and greenhouse expansion. FAO GIEWS NDVI and Vegetation Health Index products were incorporated only as regional environmental context. The resulting spatial database provides a robust framework for sustainable pasture planning, grazing management, and future land-use policy development in South Lebanon.

1. Introduction

Sustainable Pasture Systems (SPSs) integrate livestock, pasture resources, woody vegetation, and grazing management within a multifunctional land-use framework. They are increasingly recognized as a strategy for reconciling livestock production with the conservation of ecosystem services, including soil protection, water regulation, carbon storage, biodiversity conservation, and rural livelihood support. In contrast to conventional grazing systems that often emphasize short-term forage use, SPSs require an explicitly management-oriented approach in which vegetation condition, grazing pressure, land-use change, and institutional capacity are evaluated together [1,2,3,4].

Silvopastoral systems represent one of the most widely studied forms of SPS. By combining trees, forage species, and livestock within the same management unit, these systems can moderate microclimate, reduce heat stress, stabilize soils, improve nutrient cycling, and diversify farm outputs. Their performance, however, depends strongly on on-site conditions, species selection, grazing intensity, farmer participation, and long-term monitoring. As a result, an SPS assessment cannot be limited to detecting vegetation cover alone; it must also consider whether spatial patterns can inform practical management decisions [5,6,7,8,9].

Existing SPS assessment methods generally fall into five complementary groups: field-based pasture and biomass measurements, livestock productivity indicators, biodiversity and eco-system-service assessments, socio-economic and governance evaluations, and remote-sensing or GIS-based spatial monitoring. Field methods provide detailed information on forage quantity, quality, species composition, and carrying capacity, but they are costly and difficult to apply consistently over large areas or in insecure regions. Ecosystem-service and socio-economic assessments are highly relevant for planning, but they require extensive local data. Remote sensing, by contrast, offers repeated spatial observations over long periods and is particularly useful where field access is limited; however, it provides indirect indicators such as greenness and stress rather than direct measurements of forage nutritive value or grazing capacity [10,11,12,13].

This distinction is central to the present study. NDVI and VHI are not treated as direct measures of SPS performance. Instead, they are used as spatial indicators of vegetation condition, seasonal stress, and broad degradation or recovery tendencies that can support SPS planning. This interpretation follows recent work in Mediterranean silvopastoral systems showing that NDVI can contribute to pasture-quality assessment when supported by field data while also requiring caution because vegetation greenness does not necessarily represent usable biomass, crude protein, neutral detergent fiber, or livestock carrying capacity [11,12].

In Mediterranean and semi-arid environments, pasture systems are shaped by interacting with ecological, climatic, and socio-economic pressures. Drought, irregular rainfall, overgrazing, land abandonment, urban expansion, wildfire, and fragmented governance can all modify vegetation structure. These pressures are particularly important in Lebanon, where traditional pastoral and agro-pastoral systems have been disrupted by rapid land-use change, limited institutional support, and recurrent conflict.

The Lebanese case therefore requires a socio-ecological interpretation of pasture sustainability. SPS planning is not only a technical question of where vegetation exists but also a question of resilience, resource governance, and land-use trade-offs. Resilience refers to the capacity of pasture landscapes and pastoral communities to absorb disturbance while maintaining essential ecological and livelihood functions [14,15,16,17,18,19]. Governance is critical because many grazing resources function as shared or semi-shared resources, requiring coordination among herders, landowners, municipalities, and public institutions [19]. Land-use trade-offs are equally important because pasturelands compete with agriculture, settlement expansion, forestry, and conservation priorities [20,21,22].

Against this background, the present study addresses a specific research gap: the absence of long-term, spatially explicit evidence that connects grassland dynamics in Lebanon with management-oriented SPS planning. This research maps land cover from 2000 to 2024, with a focus on South Lebanon and Nabatieh, and analyzes grassland for recent vegetation-stress interpretation. The objective is not to claim a complete suitability analysis of SPS implementation but to establish a spatial baseline that identifies grassland trends, highlights stress patterns, and supports targeted rehabilitation and monitoring decisions in a Mediterranean, conflict-affected context.

2. Study Area

Lebanon is characterized by rangelands that occur across a steep Mediterranean environmental gradient, from humid coastal and mountain zones to semi-arid inland areas. This diversity allows long-term grassland dynamics to be interpreted across contrasting ecological settings while maintaining a single national land-cover dataset [23,24,25,26].

South Lebanon was selected as the regional focus because it combines ecological im-portance, dependence on natural grazing resources, recurrent vegetation stress, and strong land-use and conflict-related pressures. The region includes the South and Nabatieh governorates and extends from the Mediterranean coastal plain to inland hilly and mountainous landscapes. This spatial heterogeneity makes it a relevant case for examining how grassland monitoring can support SPS planning under Mediterranean and post-conflict conditions [22,27,28,29,30].

The study area covers the South Lebanon and Nabatieh governorates for the period of 2000–2024. This provides recent contextual evidence of seasonal vegetation stress in the priority SPS planning area.

The South Lebanon and Nabatieh governorates are representative of many Mediterranean pasture landscapes because grazing resources are distributed across fragmented agricultural, shrubland, forest-edge, and semi-natural grassland mosaics (Figure 1). Pasture availability depends strongly on winter–spring rainfall, while summer drought limits biomass production. These conditions, combined with land-use pressure and conflict-related access restrictions, justify the selection of the region as a critical case study for management-oriented SPS assessment [23,24,25].

The study-area description was intentionally limited to characteristics directly relevant to pasture monitoring: topography, Mediterranean seasonality, grazing dependence, land-cover heterogeneity, and access constraints. Broader historical and socio-economic details were reduced to maintain focus on the spatial assessment and management interpretation (Figure 2).

The rangelands and forest-edge pastures of South Lebanon are important for livestock support, biodiversity conservation, and rural livelihoods, but they remain vulnerable to overgrazing, urban expansion, wildfire, drought, and disturbance caused by conflict. These pressures can alter vegetation structure and complicate the interpretation of satellite-derived recovery signals [15,24,25,26,27,28,29].

Because direct field access was restricted during the study period, the regional analysis relied on satellite-based indicators and secondary spatial information. This limitation does not invalidate the use of remote sensing for broad monitoring, but it prevents the study from making direct claims about forage quality, carrying capacity, or livestock productivity without complementary field data [10,11,12,13].

Overall, South Lebanon provides a relevant case for assessing how GIS and vegetation indices can support SPS planning where environmental pressures and access limitations make systematic field surveys difficult [10,11,12,13,15,29].

3. Materials and Methods

The workflow was organized into two complementary analytical levels (

Table 1. Comparison of the main spatial datasets and indicators used in the study.

Component	Data Source/Method	Spatial Scale	Temporal Coverage	Main Purpose	Interpretation Limits
Land-cover classification	Landsat ETM+, as well as Landstat 8 and 9 OLI imagery, classified using Random Forest	South Lebanon and Nabatiyeh provinces	2000–2024	Mapping dense and dispersed grassland dynamics	Requires independent accuracy assessment; not a direct measure of forage quality or carrying capacity
Dense and dispersed grassland area	Extracted from classified land-cover maps	South Lebanon and Nabatiyeh provinces	2000–2024	Quantifying long-term grassland change	Does not provide species composition, biomass, or grazing intensity
NDVI profiles	FAO GIEWS Earth Observation products	Crop-area mask of South Lebanon and Nabatieh provinces	2022–2024	Contextual interpretation of vegetation greenness and seasonality	Crop-area based; not pasture-specific validation
VHI profiles	FAO GIEWS Earth Observation products	Crop-area mask of South Lebanon and Nabatieh	2022–2024	Contextual interpretation of drought-related vegetation stress	Indicates vegetation stress but not forage quality, productivity, or stocking capacity
Management interpretation	Integration of land-cover change, NDVI/VHI patterns, and literature	South Lebanon planning context	2000–2024 and 2022–2024	Identifying priority zones for monitoring, restoration, and SPS planning	Requires field validation, herder knowledge, tenure information, and local governance assessment

).

First, Landsat imagery was used for land-cover analysis from 2000 to 2024 to map the major land-cover classes and distinguish pasture from other land-cover classes, as well as dense from dispersed grasslands. Second, regional NDVI and VHI profiles were interpreted for South Lebanon and Nabatieh for 2022–2024 as contextual evidence of recent vegetation stress. These two analytical levels were not treated as identical datasets: the Landsat classification provides the basis for grassland change analysis, whereas the regional vegetation-index profiles provide supplementary seasonal information for the planning context [10,11,12,13].

3.1. Data Acquisition

Landsat Collection 2 Level-2 Surface Reflectance (SR) products were used to analyze grassland dynamics in South Lebanon between 2000 and 2024. Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI, and Landsat 9 OLI-2 imagery was obtained through Google Earth Engine (GEE). Surface reflectance products were selected because they are atmospherically corrected and provide consistent spectral information suitable for long-term land-cover monitoring. Atmospheric correction was performed using the Landsat Ecosystem Disturbance Adaptive Processing System (LEDAPS) for Landsat 5 and Landsat 7 imagery and the Land Surface Reflectance Code (LaSRC) for Landsat 8 and Landsat 9 imagery. All images were projected to the WGS84/UTM Zone 36N coordinate system. Cloud-contaminated pixels were removed using the Quality Assessment (QA_PIXEL) band available within Collection 2 products. For Landsat 7 imagery acquired after the Scan Line Corrector (SLC) failure in 2003, image mosaicking and temporal compositing were used to reduce data gaps. Representative cloud-free images acquired during April and May were selected because these months correspond to the period of maximum herbaceous vegetation development in Mediterranean ecosystems while minimizing seasonal variability among years.

Three Landsat collections are defined:

• L7 (2000–2013): Landsat 7 ETM+;
• L8 (2014–2021): Landsat 8;
• L9 (2022–2024): Landsat 9.

Each collection was filtered by path/row (174/36 and 174/37) and by cloud cover (<1%).

The image collections were restricted to April–May, when natural pastures and rangelands in Lebanon typically reach maximum spring greenness.

Seventy-three samples were obtained from point shapefiles using Google Earth imagery and expert interpretation.

3.2. Correction of the Images and Classification

Land-cover classification was performed using the Random Forest (RF) algorithm implemented within the Google Earth Engine. Random Forest is a non-parametric ensemble machine-learning classifier that constructs multiple decision trees using bootstrap sampling and random feature selection, providing high classification accuracy for heterogeneous landscapes. Predictor variables included surface reflectance values from all available spectral bands together with vegetation indices. Training samples representing the principal land-cover classes were manually digitized using high-resolution Google Earth imagery and expert interpretation. The classifier was trained separately for each observation year to account for interannual spectral variability.

Preprocessing:

All Landsat TM, ETM+, and OLI images were obtained as Level 2 surface reflectance products from the USGS archive. These products include atmospheric correction using the LEDAPS algorithm for TM/ETM+ and LaSRC for OLI, ensuring radiometric consistency across sensors [31,32,33]. Standard geometric correction was applied to align imagery to the WGS84/UTM projection.

Sensor harmonization:

To address spectral differences between TM/ETM+ and OLI, we applied the published cross sensor harmonization coefficients developed in [31], which adjust OLI reflectance to be consistent with TM/ETM+. This step reduces spurious interannual variability caused by sensor-response differences.

ET+ SLC-off gap:

We acknowledge the Scan Line Corrector (SLC) failure in Landsat 7 ETM+ after May 2003. To mitigate striping and missing data, we used gap-filling procedures. Gap-filling for Landsat 7 SLC-off data was performed using focal mean interpolation with a 3 × 3 kernel, followed by temporal mosaicking across adjacent paths/rows. This configuration minimized striping artifacts while retaining >99% coverage of the study area.

Sampling design:

Training samples were compiled using a combination of high-resolution Google Earth imagery, existing land-cover maps, and recent field observations. A total of 73 points were collected, and these points were distributed across all ten land-cover classes: agriculture, forest, pasture, greenhouse, road, bare rock, sand, bare land, water body, and urban. The sampling density for each class varied depending on the availability and reliability of reference information in different parts of South Lebanon and Nabatieh. Dense grasslands were defined as polygons with ≥70% herbaceous pixel fraction and NDVI ≥ 0.6, whereas dispersed grasslands were defined as polygons with <70% herbaceous cover and NDVI values between 0.2 and 0.6. This quantitative rule ensures reproducibility and consistent class separation across the study area. In addition, independent validation datasets were generated separately for each year, ensuring that classification accuracy was assessed reliably and without overlap with the training samples.

Classification accuracy was evaluated using an independent validation dataset. Validation samples were randomly distributed across the study area and compared with reference information obtained from high-resolution imagery. Confusion matrices were generated for each classified image to calculate the Overall Accuracy (OA), Producer’s Accuracy (PA), User’s Accuracy (UA), and Kappa coefficient. These metrics were used to assess the reliability of the classification results and ensure consistency among observation years.

3.3. Land-Cover Classification

Supervised land-cover classification was performed using the Random Forest (RF) algorithm [14]. RF was chosen because it is a non-parametric ensemble method that performs well with heterogeneous landscapes, nonlinear spectral responses, and multi-band satellite data. The algorithm reduces the risk of overfitting by combining multiple decision trees and is widely used in remote-sensing land-cover classification. The classification distinguished dense grasslands, dispersed grasslands, forests, agricultural land, urban areas, bare surfaces, and water bodies.

The Random Forest model was optimized to balance classification accuracy, generalization, and computational efficiency. The number of trees increased to 300, which improved stability and reduced variance compared with the initial 12-tree setup. The square root of the total number of predictors was used for the maximum-features parameter in classification tasks to promote diversity among trees and reduce overfitting. The minimum number of samples per split was set to 5, and the minimum number of samples per leaf was set to 2. Bootstrap sampling was retained to improve robustness through resampling, and class weights were set to “balanced” for imbalanced datasets to improve the representation of minority classes. Different split criteria, including the Gini coefficient and entropy for classification and mean squared error for regression, were evaluated to identify the most consistent configuration. Together, these parameter settings improved model performance and generalization across datasets.

The pixel distribution across classes was recorded as follows: agriculture, 19.2%; forest, 8.2%; pasture, 21.9%; greenhouse, 5.5%; road, 12.3%; bare rock, 6.8%; sand, 9.6%; bare land, 6.8%; water body, 1.4%; and urban, 8.2%. This revealed a clear imbalance, with pasture and agriculture lands dominating the dataset while classes such as water body and greenhouse were under-represented. To address this imbalance, class weights were set to ‘balanced’ in the Random Forest, ensuring minority classes such as dispersed pasture and water body were adequately represented. Validation confirmed that balanced weighting improved the producer’s accuracy for under-represented classes.

Feature importance ranking was computed from the Random Forest model to identify the most influential spectral bands and vegetation indices. Results consistently showed NDVI and SWIR bands as the strongest predictors, followed by red and NIR reflectance.

3.4. Validation of Classification Results Using Confusion Matrix, Overall Accuracy, Producer’s Accuracy, User’s Accuracy, and Kappa Coefficient

The reliability of land-cover classification results must be assessed using robust statistical measures that go beyond visual inspection. A widely accepted approach is the use of a confusion matrix, which provides a pixel-by-pixel comparison between reference data and classified outputs [32]. A confusion matrix is a contingency table where rows represent the true classes and columns represent the predicted classes. From this matrix, four key indices are derived: Overall Accuracy (OA), the Producer’s Accuracy (PA), the User’s Accuracy (UA), and the Kappa coefficient (Κ) [33]. These metrics are essential for validating classification performance and ensuring that spatial claims are statistically supported.

Overall Accuracy (OA) is calculated as the proportion of correctly classified samples relative to the total number of samples, expressed mathematically as follows:

$$OA= {\displaystyle \frac{{\displaystyle {\sum }_{i=1}^m}{n}_{ii}}{N}}$$

(1)

$${PA}_i={\displaystyle \frac{\mathrm{C}\mathrm{o}\mathrm{r}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{t}\mathrm{l}\mathrm{y} \mathrm{c}\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{s}\mathrm{i}\mathrm{f}\mathrm{i}\mathrm{e}\mathrm{d} \mathrm{s}\mathrm{a}\mathrm{m}\mathrm{p}\mathrm{l}\mathrm{e}\mathrm{s} \mathrm{o}\mathrm{f} \mathrm{c}\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{s} \mathrm{i}}{\mathrm{T}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l} \mathrm{a}\mathrm{c}\mathrm{t}\mathrm{u}\mathrm{a}\mathrm{l} \mathrm{s}\mathrm{a}\mathrm{m}\mathrm{p}\mathrm{l}\mathrm{e}\mathrm{s} \mathrm{o}\mathrm{f} \mathrm{c}\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{s} \mathrm{i}}}$$

(2)

$${UA}_i={\displaystyle \frac{\mathrm{C}\mathrm{o}\mathrm{r}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{t}\mathrm{l}\mathrm{y} \mathrm{c}\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{s}\mathrm{i}\mathrm{f}\mathrm{i}\mathrm{e}\mathrm{d} \mathrm{s}\mathrm{a}\mathrm{m}\mathrm{p}\mathrm{l}\mathrm{e}\mathrm{s} \mathrm{o}\mathrm{f} \mathrm{c}\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{s} \mathrm{i}}{\mathrm{T}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l} \mathrm{p}\mathrm{r}\mathrm{e}\mathrm{d}\mathrm{i}\mathrm{c}\mathrm{t}\mathrm{e}\mathrm{d} \mathrm{s}\mathrm{a}\mathrm{m}\mathrm{p}\mathrm{l}\mathrm{e}\mathrm{s} \mathrm{o}\mathrm{f} \mathrm{c}\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{s} \mathrm{i}}}$$

(3)

To address this limitation, the Kappa index (Κ) is used. Kappa measures the degree of agreement between classification and reference data after correcting for chance agreement. It is defined as

$$\mathrm{K}={\displaystyle \frac{OA-P_e}{1-P_e}},$$

(4)

where $n_{ii}$ is the number of correctly classified samples for class i (diagonal of the confusion matrix), m is number of classes, N is the total number of samples, and $P_e$ = expected agreement by chance= Compute row totals × column totals ÷ total².

3.5. Regional Vegetation-Stress Analysis (South Lebanon and Nabatieh, 2022–2024)

Vegetation condition and stress dynamics in South Lebanon and Nabatieh were evaluated using the Normalized Difference Vegetation Index (NDVI) and Vegetation Health Index (VHI). The NDVI was interpreted as an indicator of vegetation greenness and photosynthetic activity, while the VHI was interpreted as a drought-related vegetation-stress indicator combining vegetation condition and thermal stress. In this study, both indices were used cautiously as indicators of vegetation condition and stress; they were not interpreted as direct measurements of forage quality, biomass availability, species composition, nutritional value, or livestock carrying capacity [10,11,13,15,18].

The NDVI and VHI profiles for South Lebanon and Nabatieh were obtained from FAO GIEWS Earth Observation products. These profiles are based on FAO’s crop-area mask and therefore describe vegetation behavior within the agricultural/crop-area domain rather than pasture-specific pixels. For this reason, the regional NDVI/VHI figures were retained only as contextual evidence of recent seasonal vegetation stress in the wider agricultural landscape of South Lebanon. They were not used to validate the Landsat grassland classification and are not presented as direct measurements of pasture condition.

This distinction responds to an important methodological limitation. The Landsat classification is the primary source for the long-term grassland-change record, whereas the regional NDVI/VHI profiles provide supplementary temporal information on recent greenness and drought stress. The use of these profiles is justified because SPS planning in South Lebanon is embedded within mixed agricultural and grazing landscapes; however, pasture-specific NDVI/VHI extraction should be developed in future work using a pasture mask derived from validated land-cover data [10,11,12,13].

The combined framework therefore links long-term grassland mapping with recent vegetation-stress context while explicitly separating what is directly measured from what is inferred for management. This approach supports SPS planning by identifying broad areas of grassland change and recurrent seasonal stress, but it does not replace field-based assessment of forage quality or stocking capacity.

Overall, the integration of land-cover classification with vegetation index analysis provides a management-oriented framework for identifying grassland change, recurrent vegetation stress, and priority areas for SPS monitoring and rehabilitation under changing climatic, land-use, and access conditions [10,11,12,13,19,29].

3.6. NDVI Anomaly Detection During Conflict Periods

To quantitatively assess the impact of conflict on grassland dynamics, NDVI time series were analyzed using the Breaks for Additive Season and Trend (BFAST) algorithm. This method decomposes the NDVI into trend, seasonal, and remainder components and identifies statistically significant breakpoints. Breakpoints detected during conflict years were interpreted as disturbance signals, allowing us to separate conflict-related anomalies from climatic variability.

4. Results

The analysis began with the classification of corrected, enhanced, and clipped Land-sat images acquired between 2000 and 2024 to generate land-cover maps for the selected years. Grassland pixels were subsequently identified as pasture areas, and the NDVI was used to distinguish dense from dispersed pasture.

4.1. Land Cover Classification

Random Forest classification was applied to the selected observation years. The resulting maps were validated using the confusion matrix, Overall Accuracy (OA), Producer’s Accuracy (PA), User’s Accuracy (UA), and Kappa coefficient as described in Section 3.4. Figure 3 presents confusion matrices for representative years, and Figure 4 shows the PA and UA for most of the covered years. In addition, Figure 5 presents selected land-cover maps.

The following charts (Figure 6) show the OA and Kappa coefficient for different years from 2000 to 2024.

4.2. Separation of Dense and Dispersed Pasture

Pasturelands are often classified as a single “grassland” or “herbaceous” category in land-cover maps. However, ecological and agricultural applications require distinguishing between dense pasture (continuous, healthy vegetation) and dispersed pasture (fragmented or sparse vegetation). The Normalized Difference Vegetation Index (NDVI) provides a quantitative measure of vegetation greenness and photosynthetic activity, making it a robust tool for this separation:

$$NDVI={\displaystyle \frac{NIR-RED}{NIR+RED}},$$

(5)

where NIR is near-infrared reflectance and RED is red reflectance. NDVI values range between −1 and +1, with higher values indicating denser vegetation cover.

Several studies have established threshold ranges for vegetation density (Figures S1–S18 in Supplementary Materials). Dense vegetation is generally associated with NDVI values above 0.6, while sparse or dispersed vegetation corresponds to values between 0.2 and 0.6. Values below 0.2 typically represent bare soil or non-vegetated surfaces [30,31]. In the context of pasture mapping, this study adopts the following thresholds:

-

Dense pasture: NDVI ≥ 0.6;

-

Dispersed pasture: 0.2 ≤ NDVI < 0.6.

The workflow involves the following:

-

Extracting grassland pixels from the supervised land-cover classification;

-

Computing the NDVI for each pixel using Landsat imagery;

-

Applying the thresholds above to subdivide grassland into dense and dispersed pasture. Figure 7 and Figure 8 show the NDVI maps and the dense and dispersed pasture maps for studied years.

The graph in Figure 9 illustrates contrasting trajectories of dispersed and dense pasture over the study period. Dispersed pasture (light green line) shows repeated peaks around 2007, 2011, and 2019, followed by a marked decline after 2020, indicating progressive fragmentation and loss of continuity. In contrast, dense pasture (dark green line) remained relatively stable until 2010, then began a gradual increase, with a sharp rise after 2022. Together, these trends highlight a long-term redistribution of pasture cover, with dispersed areas contracting while dense pasture expands, reflecting intensification and structural change in land use.

4.3. Analysis of Transitions in Land-Cover Classes Between 2000 and 2024

A pixel-by-pixel change-detection analysis was performed between the classified land-cover maps for 2000 and 2024. The resulting transition matrix (

Table 2. Transition matrix (2000 (rows) → 2024, in hectares).

Class	Agriculture	Forest	Pasture	Greenhouse	Road	Bare Rock	Sand	Bare Land	Water	Urban
Agriculture	0.00	19.50	13,800.00	9800.00	20,300.00	8400.00	24,600.00	15,500.00	145.00	404.70
Forest	0.00	0.40	420.70	3.00	1182.40	81.70	924.00	5913.80	25.50	32.00
Pasture	0.00	60.70	63,784.30	56.10	973.80	7077.00	11,231.00	157.80	91.00	279.00
Greenhouse	0.00	0.00	84.10	852.00	181.50	58.90	259.60	114.20	0.40	1.70
Road	0.10	1.40	418.00	875.00	90.20	144.20	104.60	38.50	4.50	9.60
Bare rock	0.00	0.00	1330.60	147.80	543.00	401.00	492.00	128.70	4.00	13.80
Sand	0.00	0.00	1004.90	79.30	204.30	144.20	277.60	62.50	1.70	2.50
Bare land	0.00	2.40	2199.60	220.00	202.00	427.00	268.00	61.90	11.40	30.50
Water	0.20	25.50	210.40	369.00	3.50	2.50	3.50	7.80	6.00	0.60
Urban	0.00	10.30	4310.00	0.00	167.00	0.00	0.00	0.00	0.00	460.80

) quantifies persistence, gains, losses, and conversions among land-cover classes, with all values expressed in hectares.

The matrix reveals substantial transformations across the landscape over the 24-year period. Pasture, which covered a large portion of the study area in 2000, experienced extensive conversion. Only 63,784 ha remained as pasture in 2024, while notable portions transitioned to greenhouse areas (56 ha), bare rock (7077 ha), sand (11,231 ha), and urban land (279 ha). These transitions highlight the progressive fragmentation and repurposing of rangeland toward agricultural intensification and built-up development.

Agricultural land also underwent considerable change. Approximately 13,800 ha shifted from agriculture to pasture, while 9800 ha transitioned to greenhouse areas. Additional conversions from agriculture to road surfaces (20,300 ha), sand (24,600 ha), and bare land (15,500 ha) reflect the expansion of infrastructure, quarrying, and soil-exposed surfaces.

Forest areas remained relatively stable, with small transitions toward pasture (420.7 ha) and greenhouse (3 ha). Greenhouse areas showed moderate persistence (852 ha) and modest gains from pasture, agriculture, and forest. Urban expansion was primarily fueled by conversions from pasture (4310 ha) and agriculture (404.7 ha), consistent with observed settlement growth.

Overall, the transition matrix illustrates a clear trend toward intensified agriculture, greenhouse expansion, and urban development, accompanied by reductions in natural and semi-natural land-cover types such as pasture and forest.

4.4. NDVI-Based Discrimination of Grassland Density

NDVI analysis was used to compare the spectral behavior of grasslands with that of other dominant land-cover types across Lebanon for 2000–2024. NDVI values ranged between −1 and +1, with higher positive values generally indicating dense green vegetation, values close to zero indicating bare or built surfaces, and negative values commonly associated with water. Because the comparison includes heterogeneous land-cover classes and multi-date imagery, the values should be interpreted as class-level spectral tendencies rather than precise biophysical measurements.

Agricultural lands exhibited consistently higher NDVI values, generally ranging from 0.3 to 0.7, reflecting seasonal crop cycles and varying management practices. Urban areas showed consistently low NDVI values of approximately 0.15, indicating limited vegetative cover. Water bodies displayed negative NDVI values, while bare rocky surfaces remained close to zero throughout the study period.

In contrast, grasslands exhibited relatively stable NDVI values with limited interannual fluctuation compared to agricultural areas (Figure 10). This stability reflects the perennial nature of grassland vegetation and its relative insensitivity to short-term land-use changes. Based on these NDVI characteristics, grasslands were successfully separated from other land-cover classes and subsequently classified into dense and dispersed grasslands.

4.5. Vegetation Stress in South Lebanon: NDVI and VHI Profiles (2022–2024)

Seasonal vegetation dynamics in South Lebanon and Nabatieh were examined using FAO GIEWS NDVI and VHI profiles for 2022–2024. These profiles represent crop-area vegetation behavior and are used here as contextual evidence of seasonal stress in the wider agricultural landscape. The NDVI increased during spring and declined sharply during the summer months (Figure 11 and Figure 12).

In Nabatieh, NDVI values peaked at approximately 0.60 in 2022 and 2023 and approximately 0.65 in 2024. All three years showed pronounced summer declines, with the strongest decline observed in 2023, when the NDVI fell below the long-term average.

VHI profiles (Figure 13 and Figure 14) showed a comparable seasonal pattern. Both South Lebanon and Nabatieh exhibited high VHI values during spring, followed by sharp summer declines. Values fell below 0.4 in 2023, while 2024 showed improved early-season vegetation health followed by persistent summer decline.

4.6. Contextual Reference Information and Conflict-Related Impacts

Historical information from the Ministry of Agriculture and CNRS provided useful contextual material for interpreting the distribution of grazing lands in South Lebanon. However, these sources do not constitute a formal accuracy assessment of the present classification. They were therefore used only to support the qualitative interpretation of grassland occurrence and land-use context.

5. Discussion

5.1. General Findings and Considerations

The present study provides a long-term spatial assessment of grassland dynamics and a regional interpretation of recent vegetation stress in South Lebanon and Nabatieh. The Landsat-based classification revealed substantial temporal and spatial changes in dense and dispersed grasslands between 2000 and 2024, while the pixel-based transition analysis provided quantitative evidence of persistence and conversion among the principal land-cover classes. The FAO GIEWS NDVI and VHI profiles complemented these results by providing information on recent seasonal vegetation stress within the wider crop-area landscape of South Lebanon and Nabatieh. These datasets should nevertheless be interpreted differently: the Landsat classification and transition analysis provide the primary evidence of long-term land-cover change, whereas the FAO GIEWS products provide regional environmental context rather than pasture-specific validation. This distinction is important because remote sensing can characterize vegetation greenness, cover, and stress but cannot independently determine forage quality, species composition, grazing suitability, or livestock carrying capacity [10,11,12,13,15].

The temporal analysis showed substantial variation in the extent and distribution of dense and dispersed grasslands. Dense grassland in the South and Nabatieh governorates increased from approximately 0.3 km² in 2000 to about 333 km² in 2024. Dispersed grassland occupied a considerably larger area in 2000 (about 837 km²) but declined to reach about 330 km² in 2024. NDVI anomaly detection confirmed that conflict years were associated with significant breakpoints in vegetation dynamics, reinforcing the interpretation that conflict contributed to abrupt disturbances in pasture cover. These patterns suggest that grassland structure is dynamic and potentially sensitive to interactions with climatic and anthropogenic pressures. Previous studies in Lebanon have identified overgrazing, land-use conversion, fragmentation, rainfall variability, and human disturbance as important pressures affecting rangeland condition [25,27,29]. However, the observed temporal changes should not be attributed to any single driver because field-based information on grazing intensity, biomass, vegetation composition, and management practices was not consistently available throughout the study period.

Importantly, the pixel-by-pixel transition analysis between 2000 and 2024 strengthens the interpretation of these spatial changes by quantifying persistence, gains, losses, and conversions among land-cover classes. The transition matrix revealed substantial conversion of agricultural land, particularly toward greenhouse areas, and identified important exchanges involving grassland. Approximately 82,795 ha classified as agriculture in 2000 were classified as greenhouse in 2024, representing the largest transition identified in the analysis. Approximately 12,556 ha changed from agriculture to grassland, while approximately 14,370 ha changed from grassland to greenhouse. Grassland also retained approximately 1377 ha according to the diagonal persistence values of the transition matrix. These results demonstrate that the observed landscape changes are not limited to fluctuations in total grassland area but involve substantial redistribution among agriculture, greenhouse production, grassland, and other land-cover classes.

The strong expansion of greenhouse areas, particularly at the expense of agricultural and grassland classes, indicates an important structural transformation of the landscape. Conversion from grassland to greenhouse production may represent increasing pressure on natural or semi-natural grazing resources, while transitions from agriculture to grassland may reflect land abandonment, vegetation recovery, changes in management, or other land-use processes. These interpretations should remain cautious because the transition matrix identifies where class conversions occurred but does not, by itself, establish their underlying causes. Nevertheless, the quantitative transition analysis provides a stronger basis for interpreting land-use dynamics than comparison of total class areas alone and highlights areas where future field investigation and land-management assessment should be prioritized.

The NDVI analysis further demonstrated clear spectral differences among the principal land-cover classes. Agricultural areas exhibited greater variability in the NDVI, reflecting differences in crop calendars, irrigation, harvesting, fallow periods, and management intensity, whereas urban and bare surfaces generally maintained lower vegetation-index values. Grasslands showed comparatively stable NDVI behavior, although seasonal and interannual variations remained evident. These patterns are consistent with the sensitivity of the NDVI to vegetation density, seasonal growth, and mixed-pixel effects [10,12].

The spatial redistribution of grasslands also indicates heterogeneous trajectories across the study area. Areas showing stable or increasing dense grassland cover may reflect vegetation regeneration, reduced disturbance, favorable climatic conditions, or changes in land management, whereas declining or increasingly dispersed grasslands may indicate fragmentation, conversion, or persistent disturbance. The transition analysis provides additional quantitative evidence for these landscape changes, particularly by identifying conversions between grassland, agriculture, and greenhouse production. However, determining the ecological condition of individual transition zones will require field validation and information on grazing intensity, land tenure, vegetation composition, and management history.

The regional NDVI and VHI profiles for South Lebanon and Nabatieh revealed a consistent seasonal pattern characterized by improved vegetation conditions during the spring, followed by pronounced declines during the dry summer period. In Nabatieh, the NDVI reached approximately 0.60 in 2022 and 2023 and approximately 0.65 in 2024, while stronger summer declines were observed in 2023. The VHI showed a comparable seasonal response, with values declining below 0.4 during summer 2023. These patterns indicate recurrent seasonal vegetation stress within the wider agricultural landscape and are consistent with Mediterranean climatic conditions [11,12,15,24,25]. Because the FAO GIEWS profiles are based on crop-area masks rather than pasture-specific pixels, they should be interpreted as regional contextual evidence rather than direct measurements of pasture condition.

From an SPS perspective, the combined evidence from land-cover classification, transition analysis, and vegetation-stress indicators suggests that different grassland areas require different management responses. Stable or recovering dense grasslands may primarily require protection, controlled grazing, and continued monitoring, whereas dispersed or declining grasslands may require restoration, erosion control, and regulation of grazing pressure. Areas affected by conversion to intensive agricultural uses, particularly greenhouse expansion, deserve specific attention because continued conversion could further reduce or fragment available grazing resources. Areas experiencing recurrent seasonal vegetation stress should also be considered priorities for water conservation, drought-adapted forage management, and integration of suitable native woody vegetation [3,4,6,19,22,29].

The relevance of these findings is particularly important in South Lebanon, where conflict-related access restrictions and land disturbance can alter both vegetation condition and the spatial distribution of grazing pressure. Restricted access to traditional grazing areas may concentrate livestock in accessible locations, potentially increasing localized degradation. GIS-based monitoring provides an important means of observing landscape changes under conditions in which systematic field surveys are difficult or unsafe. Nevertheless, remote-sensing observations should be combined with local knowledge, herder participation, field assessment, and institutional coordination before specific management interventions are implemented [12,16,17,18,19,28,29,30,31,32].

The findings therefore support the use of SPS as a planning framework but should not be interpreted as evidence that all mapped grasslands are immediately suitable for SPS implementation. Suitability depends on additional variables that were not fully represented in the present spatial analysis, including land tenure, accessibility, slope, soil properties, water availability, vegetation composition, livestock requirements, local management practices, and security conditions [20,21,22]. The present analysis should instead be considered a spatial baseline for identifying priority areas where more detailed SPS suitability assessment and field investigation are warranted.

Several limitations should be acknowledged. Although the land-cover classifications were evaluated using independent validation samples and accuracy metrics, uncertainty remains because historical imagery was obtained from multiple Landsat sensors and systematic field validation was constrained by accessibility. Standardized surface-reflectance products, cross-sensor harmonization, and consistent preprocessing were used to minimize these effects. In addition, the FAO GIEWS NDVI and VHI products are based on crop-area masks and therefore cannot be interpreted as pasture-specific measurements. Neither the NDVI nor the VHI provides direct information on forage quality, species composition, biomass availability, or carrying capacity. Future studies should therefore integrate the present spatial database with field biomass measurements, botanical surveys, grazing records, pasture-specific vegetation-index extraction, land-tenure information, and participatory assessments involving local herders and land managers.

Overall, the integration of long-term Landsat classification, pixel-based transition analysis, and regional vegetation-stress indicators provides a stronger basis for under-standing grassland dynamics and supporting SPS planning in Lebanon. The principal contribution of the study is not a complete SPS suitability model but a spatially explicit baseline that identifies long-term grassland trajectories; quantifies important land-cover conversions; highlights areas experiencing seasonal vegetation stress; and provides a basis for targeting future field validation, restoration, and sustainable grazing-management interventions.

5.2. Management Implications and Recommendations

The spatial results presented in Figure 11 and Figure 12 reveal clear patterns of grassland redistribution, seasonal vegetation stress, and long-term degradation that have direct implications for sustainable rangeland management in South Lebanon and Nabatieh. Dense pasture expanded primarily in the western and coastal foothill zones, while dispersed pasture contracted sharply in the eastern uplands and peri-urban belts. Seasonal NDVI profiles for 2022–2024 further show recurrent summer stress across both governorates, with Nabatieh exhibiting deeper mid-season declines. These spatial and temporal patterns provide the basis for defining priority restoration zones and SPS pilot partitioning areas, consistent with Lebanese and Mediterranean studies showing that effective rangeland management depends on environmental monitoring and adaptive grazing practices [3,4,24,29].

Priority restoration zones should be concentrated in areas where dispersed pasture has declined most severely (Figure 11), particularly the eastern Nabatieh uplands, the inland hilly terrain north of the Litani River, and the peri-urban fringes of Tyre and Sidon. These zones exhibit repeated NDVI depressions during summer (Figure 12), indicating structural degradation and moisture limitation. Restoration interventions in these areas should include reseeding with native herbaceous species, erosion-control structures on steep slopes, and regulated grazing corridors to reduce fragmentation. These spatially explicit zones represent the highest-risk areas for continued pasture loss and should be prioritized for rehabilitation and monitoring, consistent with Lebanese findings on erosion-driven rangeland degradation [19,20].

In contrast, SPS pilot partitioning zones should be established in areas where dense pasture has expanded or remained stable over time (Figure 11), particularly the western South Lebanon foothills, the coastal plain, and the mid-elevation mosaics west of Nabatieh. These areas show stronger NDVI recovery during spring and more moderate summer declines (Figure 12), indicating higher resilience and suitability for silvopastoral integration. SPS pilot zones in these regions can support rotational grazing demonstrations, integration of drought-tolerant woody species, and community-based grazing management trials. This approach aligns with silvopastoral research demonstrating the value of woody vegetation for restoring degraded grazing lands and enhancing ecosystem services [4,8,9,28].

The transition analysis further highlights the need to address land-use conversion pressures. Large areas of former grassland have transitioned to greenhouse production, especially in the coastal plain. These conversion hotspots should be incorporated into land-use planning frameworks to prevent further fragmentation of grazing resources. Conversely, areas transitioning from agriculture to grassland may offer opportunities for managed grazing or restoration, pending field verification of vegetation condition and actual forage value—an approach consistent with Mediterranean rangeland rehabilitation strategies [12,24]. Beyond spatial prioritization, several management principles remain essential.

First, multi-temporal satellite monitoring should be institutionalized to track degradation hotspots and seasonal stress patterns, especially where field access is limited. Satellite time series can effectively capture vegetation resistance to climatic stress, while NDVI-based approaches are increasingly used to monitor pasture condition and forage quality in Mediterranean production systems [12,24].

Second, controlled grazing management—rotational grazing, seasonal rest periods, and flexible stocking rates—should be strengthened to reduce pressure on dispersed pasture zones. The strong temporal variability observed in dispersed grasslands indicates that rest periods and seasonal adjustment are essential to limit repeated degradation [23]. This recommendation aligns with Lebanese small-ruminant production studies emphasizing adaptive strategies under semi-arid constraints and variable feed availability [15,16,28].

Third, water-management interventions such as spring protection, water-point distribution, and rainwater harvesting are critical in areas showing strong summer NDVI declines. These measures are particularly relevant in rainfed systems where pasture productivity is tightly controlled by seasonal moisture availability [20,22,28].

Fourth, soil conservation and erosion-control measures should be targeted to erosion-prone slopes within the identified restoration zones. Contour-based practices, vegetation buffers, and soil-improving amendments can stabilize vulnerable grazing lands and improve water retention, consistent with Lebanese rangeland erosion studies [19,20].

Fifth, post-conflict rehabilitation should be integrated into pasture planning in South Lebanon. Where grazing corridors are disrupted and access is restricted, spatial analysis can help identify alternative grazing zones and priority restoration sites [18,32]. In such areas, silvopastoral integration and the use of native drought-tolerant tree and shrub species can improve microclimatic buffering, soil stability, and long-term system resilience [1,3,5].

Finally, successful implementation of sustainable pasture systems will require community participation, extension support, and enabling policy instruments. Technical interventions alone are unlikely to succeed without herder involvement, local governance, and incentives for sustainable practices. The broader silvopastoral literature emphasizes that long-term benefits depend not only on ecological design but also on management capacity, adoption, and institutional support [15,25,26,27,29,34].

By linking management recommendations directly to the spatial distribution of pasture dynamics and NDVI seasonal behavior, we provide actionable, map-based guidance for restoration planning, SPS pilot deployment, and long-term rangeland governance in South Lebanon and Nabatieh.

6. Conclusions

This study demonstrates that the combination of multi-temporal Landsat imagery with GIS and Random Forest classification provides a reliable framework for monitoring long-term grassland dynamics in South Lebanon. The high classification accuracies obtained throughout the study period confirm the robustness of the proposed methodology, while NDVI-based separation of dense and dispersed grasslands offers a transparent and reproducible approach for pasture mapping.

The study also clarifies the role and limits of remote sensing in SPS assessment. Landsat-based classification can provide a useful baseline for the monitoring of dense and dispersed grassland change, while the NDVI and VHI can support the interpretation of vegetation greenness and stress. However, these indicators do not directly measure forage quality, species composition, carrying capacity, or livestock productivity. Future research should therefore combine satellite monitoring with systematic field validation, accuracy assessment, pasture-quality measurements, and participatory information from herders and local institutions. The approach developed here is transferable to other Mediterranean and post-conflict environments where long-term field access is limited but spatial evidence is urgently needed for rangeland rehabilitation and climate-resilient land-use planning.

By integrating land-cover mapping with vegetation health indicators, the study establishes an evidence-based baseline for rangeland monitoring and targeted intervention planning. The findings support sustainable pasture systems—particularly silvopastoral integration, rotational grazing, and soil and water conservation—as viable strategies to enhance pasture resilience, protect ecosystem services, and sustain rural livelihoods. Future work should prioritize field-based validation when conditions allow, integration of climate drivers for stronger attribution, and long-term monitoring of SPS pilot implementations to evaluate their effectiveness and scalability across Lebanon.

The transition matrix further quantified land-cover persistence and conversion pathways, providing objective evidence of the increasing pressure exerted by greenhouse expansion on traditional pasture resources. Although FAO GIEWS vegetation products contributed valuable regional environmental context, the principal findings are derived from the Landsat-based classification and change-detection analyses.