Machine learning integrated approach for modeling crop production in Bangladesh

Crop cultivation is vital to Bangladesh's economy, engaging about half of the labour population and contributing 17.5 % to the national GDP. This study presents a detailed modelling framework for agricultural productivity in Bangladesh employing advanced machine learning models: Artificial Neural Network (ANN), Random Forest (RF), and Support Vector Machine (SVM). The main objectives were to predict crop output, identify key environmental factors, and map areas with varying productivity levels. Thirteen variables encompassing meteorological, geographical, geological, and topographic factors including evapotranspiration, precipitation, soil moisture, and digital elevation model (DEM) were employed to assess the production potential for eight primary crops: Aman, Aush, Boro, Jute, Lentil, Maize, Potato, and Wheat. The Artificial Neural Network (ANN) had the highest predictive performance (average ROC-AUC = 0.93), succeeded by Random Forest (RF) at 0.90 and Support Vector Machine (SVM) at 0.87. Evapotranspiration emerged as the primary determinant, especially for water-intensive crops, whereas soil salinity had a minimal impact. Tangail (2.65 %), Naogaon (2.57 %), and Sylhet (2.56 %) had the most adaptability, but Rangamati (3.95 %) and Bandarban (3.19 %) shown the lowest appropriateness. The study introduces a novel, data-driven approach that integrates many environmental layers with machine learning models to improve production forecasts and inform sustainable agricultural practices. These findings significantly improve climate-resilient agricultural practices and informed policy-making for food security in Bangladesh.