Interpretable Machine Learning Framework for Early-Stage Potato Yield Forecasting Using Climatic Features in Bangladesh

Abstract

Abstract: Bangladesh, one of the world’s major potato-producing countries, faces significant yield variability driven by changing climatic conditions, with profound implications for agricultural productivity and food security. However, the specific impact of meteorological factors on potato production has not been thoroughly investigated. The present paper aims to provide a robust quantitative framework for assessing the relative effects of agro-climatic variables on potato yield formation. Weather variables that were statistically significantly related to the mortality count were first selected using Analysis of Variance (ANOVA) with F-regression feature ranking and Individual Weather Importance ranking through Random Forest. Pearson and Spearman’s correlation analyses were additionally performed to assess the strength and direction of these relationships, along with bivariate kernel density estimation (KDE) used to enlarge climatic optima toward yield maximization. For testing predictive robustness, climatic variability was taken into account, and three machine learning algorithms —RF (Random Forest), SVR (Support Vector Regression), and KNN (K-Nearest Neighbors) —were investigated. The RF model was the most accurate and generalizable, yielding almost perfect results (R² = 0.999) and minimal forecast errors (MAPE = 0.70%, MAE = 0.0803, RMSE = 0.1014). The developed framework enhances the understanding of climate–yield relationships and provides practical insights for science-based agricultural management and adaptation to climatic stress. These results have immediate implications for middle-income agrarian economies, such as Bangladesh, in strengthening resilience to climatic variability through sustainable food security.