Advancing Agricultural Land Suitability in Urbanized Semi-Arid Environments: Insights from Geospatial and Machine Learning Approaches

Abstract

Rising food demands are increasingly threatened by declining crop yields in urbanizing riverine regions of Southern Asia, exacerbated by erratic weather patterns. Optimizing agricultural land suitability (AgLS) offers a viable solution for sustainable agricultural productivity in such challenging environments. This study integrates remote sensing and field-based geospatial data with five machine learning (ML) algorithms—Naïve Bayes (NB), extra trees classifier (ETC), random forest (RF), K-nearest neighbors (KNN), and support vector machines (SVM)—alongside land-use/land-cover (LULC) considerations in the food-insecure Dharmapuri district, India. A grid searches optimized hyperparameters using factors such as slope, rainfall, temperature, texture, pH, electrical conductivity, organic carbon, available nitrogen, phosphorus, potassium, and calcium carbonate. The tuned ETC model showed the lowest root mean squared error (RMSE = 0.15), outperforming RF (RMSE = 0.18), NB (RMSE = 0.20), SVM (RMSE = 0.22), and KNN (RMSE = 0.23). The AgLS-ETC map identified 29.09% of the area as highly suitable (S1), 19.06% as moderately suitable (S2), 16.11% as marginally suitable (S3), 15.93% as currently unsuitable (N1), and 19.21% as permanently unsuitable (N2). By incorporating Landsat-8 derived LULC data to exclude forests, water bodies, and settlements, these suitability estimates were adjusted to 19.08% (S1), 14.45% (S2), 11.40% (S3), 10.48% (N1), and 9.58% (N2). Focusing on the ETC model, followed by land-use analysis, provides a robust framework for optimizing sustainable agricultural planning, ensuring the protection of ecological and social factors in developing countries.