Machine learning-based prediction of rice yield from rain feeding in monsoonal tropical area of Java, Indonesia

Rainfed rice cultivation in monsoonal tropical areas of Java, Indonesia, is challenged by nutrient deficiencies, unpredictable rainfall amounts, and limited agricultural investment, leading to fluctuating yields. The purpose of this study was to develop a precise rice yield prediction model using machine learning tailored to specific toposequences in Central Java. A combination of survey-based field and laboratory methods was employed, integrating climate, soil, socio-economic, and land management variables from 87 targeted sampling points. Machine learning analysis using Bayesian Neural Networks (BNN) demonstrated moderate accuracy with R2 = 0.840 and RMSE = 0.442 overall, but accuracy improved significantly when models were adjusted to elevation-specific categories, achieving R2 values up to 0.999. Lowland paddy field predictions were most influenced by available phosphorus (P), while rainfall, gender, education, and seed variety were key factors in medium-altitude zones; slope, available P, gender, and cropping patterns were dominant in highland areas. Pareto analysis supported the identification of these key yield determinants in each toposequence. The integration of BNN and Pareto approaches enabled the creation of a high-precision, location-specific yield prediction model. This work demonstrated that tailoring machine learning models to elevation-based agroecological zones enhances their performance and practical application. The findings are particularly valuable for agricultural stakeholders including policymakers, extension services, and farmers, who can leverage these predictive insights to optimise rainfed rice management practices and improve productivity under variable climatic conditions