Abstract
Soil degradation constitutes a critical challenge, particularly in arid and semi-arid regions, where its impacts are most severe. To mitigate these impacts, comprehensive strategies must be developed to rehabilitate and restore soil functionality. The interplay of physical, chemical, and biological characteristics jointly influences the evolution of soil processes and the restoration of its functions. This study aims to evaluate the soil quality, land degradation, and ecological risks using an integrated approach that combines GIS and remote sensing with machine learning (ML) models, specifically artificial neural network (ANN), random forest (RF), and decision tree (DT) methodologies. Utilizing Landsat ETM + images alongside a Digital Elevation Model (DEM), a geomorphological map was generated, revealing that the studied area consists of two distinct landscapes: floodplain and lacustrine plain. The results indicated that primary forms of soil degradation in the examined region included salinization, alkalization, compaction, and waterlogging. Remarkably, 90.61% of the study area was classified as high quality, whereas 9.39% was categorized as moderate quality. Furthermore, measurements of soil pollution showed considerable variation in the concentrations of trace elements throughout the area. The geo-accumulation index (I(geo)) revealed significant variations of heavy metals between different soil sample sites. Specifically, the samples exhibited pollution levels (I(geo) < 0) for As, Cd, and Se. In contrast, the levels of Cu, Pb, Zn, and U indicated a very high degree of pollution (I(geo) < 3). Furthermore, assessments of contamination degree (CD), potential ecological risk (PER), and pollution load index (PLI) demonstrated that all tested soil samples were found to be highly contaminated by the analyzed elements. The ANOVA results also indicated that there were no significant differences in model performance. Nevertheless, even slight enhancements in the accuracy of Soil Quality Index (SQI) predictions could result in substantial economic benefits and facilitate more effective resource allocation. The ANN model displayed even better accuracy for CD prediction, with R(2) values of 0.98 and 0.95 during calibration (Cal.) and validation (Val.), respectively. The DT model demonstrated exceptional performance in predicting PLI, attaining R² values of 0.99 and 0.97 during Cal. and Val., respectively. In particular, the DT model showed strong predictive accuracy for PER, with R² values of 0.97 and 0.95 during Cal. and Val., respectively. This study presents an innovative perspective on enhancing the integration of various techniques for a more comprehensive understanding of soil quality. Highlighting feature selection strategies, it aims to improve both model accuracy and interpretability.