Assessing soybean yield in Madhya Pradesh by using a multi-model approach

Abstract

Context: Soybean is a critical crop for global food security, but its yield estimation faces challenges due to climate variability and limited understanding of crop growth dynamics. Accurate yield prediction under rainfed and irrigated conditions is vital for sustainable agricultural planning and addressing the impacts of climate change. Objectives: This study aims to enhance soybean yield predictions for 16 districts of Madhya Pradesh by employing a combination of the DSSAT (CROPGRO-Soybean) model and machine learning (ML) models. It seeks to evaluate yield under varying climatic conditions and improve the precision of forecasts through a multi-model approach. Methods: The DSSAT model was parameterised using historical weather data, soil characteristics, and field experiments. Five ML models—Artificial Neural Network (ANN), K-Nearest Neighbour (KNN), Support Vector Machine (SVM), Random Forest (RF), and Decision Trees (DT)—were traineded and tested using 22 years of frontline demonstration experimental yield data. Simulated and observed values for key variables, such as leaf area index and grain yield, were compared. An ensemble ML model (EMLM) was created from the best-performing ML simulations to predict district-level soybean yields under rainfed and irrigated conditions. Results: The DSSAT model closely replicated observed data for critical variables, such as leaf area index, above-ground biomass, and grain yield. Similarly, machine learning (ML) models, such as ANN and SVM, demonstrated high accuracy in simulating soybean yields with higher correlation (r > 0.8), R²> 0.75, and RMSE (400kgha−1). The DSSAT model was further simulated for long-term soybean grain yield under rainfed and irrigated conditions. The best simulation from each trained ML model where RMSE loss is minimum was ensembled and used to predict the districts' rainfed and irrigated soybean yield. The average simulated yield of the DSSAT model was 2782kgha−1, and the EMLM was 2641kgha−1 under irrigated conditions. Meanwhile, under rainfed conditions, the DSSAT model average simulated yield was 2149kgha−1, and the EMLM model was 2047kgha−1. The multi-model comparison study for soybean yield simulation revealed that ANN and SVM models exhibited high correlation (r = 0.70&0.95) and low RMSE values (120&250kgha−1), while the RF model showed negative correlation and high RMSE with DSSAT model simulated yield. Climate variability analysis revealed positive correlations between mean seasonal rainfall and yield under rainfed conditions (r = 0.75 for DSSAT, r = 0.74 for EMLM), while solar radiation positively influenced irrigated yields (r = 0.86 for DSSAT, r = 0.84 for EMLM). However, minimum temperature negatively affected yields under irrigated conditions. Conclusions: The multi-model approach demonstrated that DSSAT and ML models, particularly ANN and SVM, effectively simulate soybean yields under diverse climatic conditions. Rainfall and solar radiation emerged as key positive drivers of yield, while higher minimum temperatures posed challenges under irrigated systems. Implications: These findings support improved precision in soybean yield predictions, offering valuable insights for agricultural planning and food security initiatives. The integration of crop models and ML approaches is crucial for mitigating climate risks and enhancing sustainable agricultural productivity. © 2024 Elsevier B.V.