Early crop yield prediction for agricultural drought monitoring using drought indices, remote sensing, and machine learning techniques

Abstract

The unpredictability of crop yield due to severe weather events such as drought and extreme heat continue to be a key worry. The present study evaluated six meteorological and three Landsat satellite-based vegetation drought indices from 1986 to 2019 in the drought-prone-semi-arid Saurashtra region of Gujarat (India). Cotton and groundnut crop yield prediction models were developed using multiple linear regression (multilayer perception (MLP)), artificial neural network with MLP, and random forest (RF). The models performed crop yield estimation at two timescales, i.e., 75 days after sowing and 105 days after sowing. The standardized precipitation evapotranspiration index/reconnaissance drought index among meteorological drought indices, normalized difference vegetation anomaly index/vegetation condition index, and normalized difference water index anomaly were chosen as best highest correlations with crop yields. The RF-based models were found most efficient in predicting the cotton and groundnut yield of Saurashtra with R2 ranging from 0.77 to 0.92, Nash–Sutcliffe efficiency ranging from 71 to 90%, and root-mean-square error ranging from 80 to 133 kg/ha for cotton and 299 to 453 kg/ha for groundnut. This study demonstrated the method for making several decisions based on early crop yield prediction including timely drought mitigation measures.