A Two-Stage XGBoost Approach for Mapping Arable and Non-Arable Soils under Salinity Stress in Southern Iraq

Abstract Selecting the optimal combination of remote sensing data, ground observation data, and a robust classification model can yield accurate estimates of soil salinization. This study, focusing on Dhi Qar, Iraq, presents a methodology for collecting data within a specific time window from 2013 to 2024, utilizing two models. Model A isolates soils with vegetation residues (VR) or non-photosynthetic vegetation (NPV) during the dry season, while Model B creates a final map classifying land as arable or non-arable based on soil salinity values. Model A acts as an extraction mask for Model B. Model A inputs include the Normalized Difference Vegetation Index (NDVI), Normalized Difference Moisture Index (NDMI), and Normalized Burn Ratio 2 (NBR2). Model B uses four salinity indices(salinity index(SI), salinity index 6(SI6), Normalized Difference Salinity Index (NDSI), and Brightness Index (BI)), land surface temperature (LST), rainfall, actual evapotranspiration (aET), and electrical conductivity (EC in dS/m). The results achieved an F-score of 94%, a Matthews Correlation Coefficient (MCC) of 90%, and an overall accuracy of 95%. The spatial distribution indicated that approximately 58% of the total area was arable, with a conductivity value of less than 7 dS/m. In contrast, non-arable land constitutes approximately 41% of the total area, exceeding 7 dS/m. Our findings suggest that a robust classification model, combined with appropriate data, significantly enhances classification and prediction accuracy, ultimately aiding in the identification of factors contributing to soil salinization.