Cloud based machine learning approaches for land use and cover changes prediction to support watershed management in the Gidabo watershed, Ethiopia

Abstract

Land use and land cover (LULC) changes in Ethiopia’s Gidabo watershed, driven by population growth and climate variability; threaten watershed sustainability and ecosystem services. Accurate LULC change mapping and prediction are vital for planning sustainable land management and climate adaptation. This study applied a cloud-based machine learning approach on the Google Earth Engine (GEE) platform to map historical (2003, 2011, 2019) and predict future (2024, 2030, 2050) LULC changes using the most accurate classifier among Random Forest (RF), Support Vector Machine (SVM), and Classification and Regression Trees (CART). The models incorporated multi-source predictors, including topographic, climatic, and socio-economic variables. The RF classifier significantly outperformed the other classifiers, achieving the highest accuracy for both historical classification (92.1% overall accuracy, Kappa 0.90) and future prediction (97.3% overall accuracy, Kappa 0.96), although some uncertainty remains in transitional classes. Historical analysis revealed a dramatic expansion of agricultural land and an increase in water bodies, largely at the expense of mixed vegetation. Predictions to 2030 and 2050 indicate a continued trend of agricultural and urban expansion, accompanied by a notable decline in mixed vegetation cover. The integration of the RF algorithm with the GEE cloud platform provides an effective and accurate approach for LULC change analysis and prediction in the Gidabo watershed. The findings deliver critical spatiotemporal insights that can directly inform sustainable land-use planning and climate-resilient watershed management strategies in the Gidabo watershed and other data-scarce regions.