Logo: to the web site of Uppsala University

Accurate drought impact-based forecasting of crop yield in India remains challenging due to the country’s hydro-climatic diversity and complex interactions between climate variability, ecosystem vulnerability, and agriculture. This study develops a framework integrating observed and forecasted drought indices across multiple accumulation periods to predict standardised crop yields at seasonal lead time before planting. Using district-level and cluster-based approaches, we apply Random Forest, Extreme Gradient Boosting, and Artificial Neural Networks to establish indicator–impact relationships for paddy rice (wet season) and wheat (dry season), leveraging historical yield data and seasonal forecasts.

District-level models outperform cluster-based ones, with Random Forest showing the best performance. Over 80% of wheat districts and 70% of rice districts achieve strong predictive accuracy, defined as RMSE below 0.2 in the test set. Incorporating ECMWF’s SEAS5 forecasts enables reliable rice yield predictions up to six months before the season—covering over 80% of wheat districts and 60–70% of rice districts. Forecast skill assessed using Continuous Ranked Probability Score (CRPS) confirms robustness across space and time, especially in districts with moderate yield variability. Weighted CRPS shows forecasts for extremely low yields (below the 10th percentile) are accurate and reliable—crucial for early warning and preparedness.

This work advances operational impact-based drought forecasting in India, offering a tool to inform anticipatory action among farmers, water managers, and supply chains. By linking drought observations and seasonal forecasts to crop yield outcomes, the study provides a replicable early warning approach to support targeted mitigation and enhance climate resilience in agriculture.