COMPARISON OF STATISTICAL AND MACHINE LEARNING-BASED APPROACHES TO ESTIMATE YIELD OF WHEAT IN THE UPPER GANGETIC PLAINS OF INDIA

Abstract

This study developed weather indices based on composite variables to analyse how multiple climate factors collectively influence the yield of wheat, which offers deeper insight into how weather variables interact to impact the growth of crop. Using historical weather data from the Indian Meteorological Department (IMD) and crop yield data from state agriculture departments, the study performed the calibration for period (2000–2017) and a validation for (2018–2019), aiming to forecast 2021 wheat yields. The SMLR model, a traditional statistical tool, was tested for normality with the Shapiro–Wilk test to ensure the compatibility with model assumptions. While most district yield datasets were normally distributed (p-value > 0.05), the Rampur and Moradabad datasets deviated (p-values of 0.005 and 0.0006, respectively). Results indicated that the ANN model outperformed SMLR, with ANN showing a mean yield deviation of 8.42% compared to 15.51% for SMLR. This superior performance is mainly due to ANN’s ability to manage collinearity among weather variables, which significantly influences yield outcomes. To assess district-level prediction accuracy, the Mean Absolute Error (MAE) was calculated for ANN, which gives result in a range from 28.84 to 212.56, with a mean of 61.78. Bulandshahar had the lowest MAE (28.84), indicating high prediction accuracy, while Saharanpur had the highest (212.56), suggesting that prediction accuracy varies by district. This variation highlights how certain regions might benefit more from the ANN model, possibly due to unique weather conditions or local factors.