Field Spectroscopy-Assisted Machine Learning Framework for Assessing Agroclimatic Effects on Paddy

Abstract

Assessment of yield-limiting factors in paddy production should be carried out properly in order to improve productivity under varying agro-climatic conditions. This research paper introduces a combined geospatial and machine learning system to estimate stage-based paddy yield stress during Kuruvai 2023 in the Thanjavur region using vegetation indices and agrometeorological variables. Five different satellite-derived vegetation indices were calibrated using spectroradiometer data during the vegetative, reproductive, and ripening stages. The field- and satellite-derived indices showed good agreement in calibration (R²=0.85-0.92). Correlation analysis showed that there were significant associations between rainfall_vegetation, humidity_reproductive, and yield variability, whereas the canopy vigor represented by ripening_EVI, SAVI correlated positively with grain filling and final productivity. Yield prediction before and after calibration was performed using machine learning models such as Multiple Linear Regression (MLR), K-Nearest Neighbors (KNN), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost). Among these, XGBoost resulted in high performance after calibration (R²=0.908, RMSE=179.728 kg/ha, MAE=143.487 kg/ha). Feature importance revealed that reproductive-stage humidity, rainfall, and ripening-stage soil moisture were the most important yield drivers. This model can be used to provide a powerful decision-support tool for spatial stress mapping, accurate irrigation regulation, and climate-resilient crop strategies in irrigated rice environments.