Crop Yield Forecasting Under Changing Climatic Conditions

Abstract

Agricultural planning, food-security management and decision making for climate-resilient agriculture require climate-sensitive crop-yield forecasting. In this research, machine learning models were built and tested for predicting the yield of various crops for different states in India as the climate changes. The analysis combined 19,652 observations from 55 crops, 30 states and union territories, six growing seasons and the time frame 1997-2019. The climatic predictors were temperature, rainfall and humidity, and the contextual variables were soil nitrogen, soil phosphorus, soil potassium, soil pH, cultivated area, crop type, season, state and year. Chronological training/validation/testing partitions for eight forecasting approaches were compared. Significant positive rainfall and national average humidity trends were observed, while there was no significant national average temperature trend. XGBoost had the lowest testing log-RMSE of 0.323 and the highest (R2) of 0.926. The most influential were crop type, area cultivated, season, soil nutrients and soil temperature. The accuracy of the forecasts was mixed between the crops and regions rice and wheat, pulses and oilseeds were comparatively good while maize forecasts were weak. The climate sensitivity analysis showed that with increased warming and a decreased level of rainfall stress or humidity, yields were predicted to decrease. The results show that the use of nonlinear machine learning models useful for agriculture planning that takes into account climate conditions however, higher resolution weather and soil-moisture and remote-sensing data required to generate more accurate future predictions.