Harnessing multi-source data and machine learning for enhanced rice yield estimation
Puja Saha, Amitabha Nath
Abstract
Rice is a major food crop of Meghalaya, a northeastern state of India, and serves as a dietary staple for nearly the entire population of the region. Currently, the state’s rice yield falls significantly short of meeting actual demand, leading to a heavy reliance on imports from other states. Unfortunately, not much research has been done on the historical trends of rice yield and its determinants. This study intends to determine the trend in rice yield by using the Mann-Kendall (MK) test and to find the joint effect of remote sensing data, climatic variables, and soil properties. The use of sophisticated machine learning techniques is to help in the understanding of these determinants and to elevate the prediction accuracy. A multi-stage method is suggested to ascertain the trend of rice yield and factors that affect it. The increase in the yield of rice and the factors affecting it are determined by using the Mann-Kendall (MK) test for the trend, and three recurrent neural network (RNN) architectures, namely Bi-LSTM, LSTM, and GRU, are used for estimating yield prediction. Historical spatiotemporal data of 70 paddy fields across the state were collected, including NDVI (Normalized Difference Vegetation Index), temperature, precipitation, wind components and various soil properties like texture, bulk density, land surface temperature and pH to build the model. Mann-Kendall (MK) test showed a statistically significant increasing trend of rice yield over the study period. Among the machine learning models evaluated, the Bi-LSTM model performed the best with an MSE of 0.0027 and MAE of 0.02407. Feature importance analysis showed that temperature, precipitation and soil properties like pH, bulk density and land surface temperature were the most important determinants of rice yield. The proposed model outperformed traditional approaches, offering a robust framework for yield prediction and decision-making in agricultural planning.