Spatiotemporal Dynamics of Drought and Desertification in Northern Algeria: A Multi-Scale Analysis Using Remote Sensing and Ensemble Machine Learning, From Regional Aridification to Local Land-Cover Transformation in the Algerian Steppe.

Abstract

Land degradation and desertification represent some of the most pressing environmental challenges facing dryland regions in the twenty-first century, where fragile ecosystems are increasingly exposed to the combined effects of climate variability and anthropogenic pressure. Northern Algeria occupies a critical transitional zone between the Mediterranean Sea and the Sahara Desert and has experienced progressive environmental degradation, reflected in declining vegetation productivity, recurrent drought, and increasing desertification severity. Despite this vulnerability, relatively few studies have examined these interconnected processes through integrated analytical approaches across multiple spatial scales. This dissertation investigates the spatiotemporal dynamics of drought, land degradation, and desertification across Northern Algeria using an integrated multi-scale framework that combines remote sensing, climatic analysis, land use/land cover (LULC) change detection, and machine learning modeling. The research is structured across three nested spatial scales: the regional scale of Northern Algeria (approximately 375,815 km²), the sub-regional scale of Northwestern Algeria, and the local scale of the Green Dam region in Djelfa Province. The analysis draws on multiple satellite and environmental datasets, including MODIS (2002–2022), Landsat (1972–2019), CHIRPS precipitation (1990–2022), ERA5 climate reanalysis (2003–2023), and SRTM DEM data, which were harmonized to a common spatial resolution of 500 m where required for the integrated analyses. Vegetation dynamics were examined using NDVI anomalies, whereas drought conditions were assessed using the Standardized Precipitation Index (SPI) at 3-, 6-, and 12-month timescales, together with the Vegetation Condition Index (VCI), Temperature Condition Index (TCI), and Vegetation Health Index (VHI). Temporal trends were evaluated using the non-parametric Mann–Kendall test and Sen’s slope estimator. To map desertification severity, a machine learning framework based on Random Forest (RF) and XGBoost was developed using 16 geo-environmental predictors and rigorously evaluated through spatial cross-validation, with 1,945 spatially stratified samples and an 80/20 training-validation split. At the local scale, supervised LULC classification and change detection were conducted for five temporal snapshots between 1972 and 2019, supported by Land Change Modeler analysis. The findings indicate a statistically significant and spatially coherent intensification of drought conditions, particularly after 2015, suggesting a shift toward more persistent hydroclimatic stress across the study region. SPI trend analysis revealed widespread drying, with 56% of the study area showing significant negative trends at SPI-3 and a mean Sen’s slope of −0.10 yr⁻¹. The vegetation-based indicators broadly reflected this deterioration, as VHI identified a substantially larger area undergoing degradation than improvement (4.89% compared with 0.51%). Correlation analysis further revealed moderate and statistically significant relationships between SPI and VHI at the short-term SPI-3 scale (r = 0.567, p = 0.007) and the seasonal SPI-6 scale (r = 0.509, p = 0.019), suggesting that vegetation health is more strongly associated with short- to medium-term precipitation variability than with longer-term moisture conditions. The year 2012 emerged as an important example of a “hot drought,” during which elevated temperatures contributed to vegetation stress despite above-average precipitation. TCI trend analysis, interpreted as an inverse temperature-based index in which declining values indicate increasing thermal stress, showed significant warming-related stress across 0.94% of the study area. At the local scale, the LULC analysis of the Green Dam region revealed a dynamic and non-linear trajectory of landscape transformation over the 47-year study period. Open plantations expanded substantially from 0 ha in 1972 to 10,345 ha in 1989, reflecting the major implementation phase of the Green Dam program, before declining to 6,303 ha by 2019, including a 43% reduction between 2009 and 2019. The 2029 scenario modeling suggests that this decline may continue, with further losses of open plantations accompanied by an expansion of degraded steppe landscapes. Population and livestock analyses also pointed to increasing anthropogenic pressure, with livestock numbers projected to exceed five million heads by 2030, a level that may place considerable pressure on the ecological carrying capacity of the steppe environment. The machine learning framework achieved high predictive performance, with an overall accuracy of 0.94 and a macro-averaged F1-score of 0.94. Feature-importance analysis consistently identified NDWI, albedo, and NDVI among the most influential predictors of desertification severity, highlighting the importance of surface moisture, soil exposure, and vegetation condition in distinguishing degradation classes. Spatial uncertainty analysis, based on the standard deviation of class probabilities, revealed relatively low uncertainty (σ < 0.15) in the northern humid areas and in zones classified as severely desertified, whereas higher uncertainty (σ > 0.25) was concentrated along ecotonal boundaries within the High Plateaus. Provincial zonal analysis identified Tlemcen, Sidi Bel Abbès, and Naâma as among the most rapidly degrading provinces, pointing to a northward shift in degradation pressure toward transitional landscapes. The illustrative 2030 projection further suggests the continued expansion of moderate, high, and severe desertification classes, particularly across the vulnerable steppe transition zone. Overall, the dissertation shows that desertification in Northern Algeria cannot be explained by climatic or anthropogenic factors in isolation. Rather, it appears to emerge from their combined and interacting effects, particularly where intensifying hydroclimatic stress coincides with sustained pressure from land-use and grazing activities. The multi-scale framework developed in this research provides an integrated basis for examining land degradation processes while offering practical tools for drought monitoring, restoration prioritization, and sustainable land management in data-scarce dryland environments. The findings also have important implications for environmental policy, supporting a shift away from rigid, uniform afforestation strategies toward restoration approaches that are more closely aligned with local ecological and hydrological conditions and that build on the resilience of native steppe ecosystems.