AI-Based Climate Pattern Analysis and Crop Recommendation for Maharashtra: A Comparative Study of Predictive Models

Abstract

This study focuses on developing and evaluating AI enabled framework for district level climate pattern analysis and crop recommendation in Maharashtra, India. The research used data from 2015 – 2024 for 36 districts covering four crops – rice, soybean, cotton, sorghum). Predictors include monthly data of IMD 0.25◦ gridded rainfall and temperature, relative humidity, solar radiation, cropping patterns, irrigation share & soil texture proxies. Three models were compared: ARIMA, LSTM and TFT. Rolling origin evaluation, ablation tests and robustness checks assessed accuracy and interpretability. Results showed clear gains from deep learning over ARIMA, with TFT achieving best generalization. Use of relative humidity and solar radiation improved error by ≈0.02–0.03 t/ha beyond rainfall–temperature inputs across crops, confirming their incremental value. The recommendation engine produced an average yield gain of ≈0.11 t/ha with ≈70% decision support accuracy. Policy implications include expanding seasonal advisories to include humidity and radiation signals, adopting risk adjusted implementation thresholds, and aligning input logistics with district level recommendations. The framework demonstrates operational feasibility for climate smart planning and provides a replicable benchmark for all states facing growing weather volatility.