Spatial Drought Prediction Using a Drought-Weighted CONVLSTM Framework with Tail-Focused Evaluation

Abstract

Accurate prediction of drought is essential to mitigate the climate risk and water resource management. The conventional methods to assess the drought predictions, emphasis on the global metrics such as the Root Mean Square, Mean Absolute Error and Correlation co-efficient and the others. This may dilute the accuracy of the prediction of the extreme events. This study introduces a Drought Weighted Convolutional Long Short-Term Memory (DW-CONVLSTM) framework for pixelwise image prediction and short-term forecasting of Standardised Precipitation Index (SPI) maps. From the monthly rainfall images, SPI3 maps are derived. Twelve-month sequences are constructed and augmented with seasonal encoding and spatial validity mask. To address the underestimation of drought severity in the deep learning methods, drought weighted masked Huber loss function is proposed to prioritise the moderate to extreme drought conditions.  The proposed framework is evaluated across four regions- Bid, Jaisalmer, Lalitpur and Tumkur. The model performance is evaluated using the global metrics such as RMSE, Correlation and R2 and tail focused conditional metrics based on the threshold values of SPI.  The proposed model DW-CONVLSTM achieves a strong predictive skill, reducing the error, while maintaining comparable global accuracy. The tail focused conditional metrics, extreme drought skill curves and spatial analysis improve spatial representation of drought severity. The results highlight the importance of drought aware learning objectives, tail focused evaluation and multi-month forecasting for reliable drought image prediction.