Persistence-Relative Evaluation of an Adaptive-Weighted LSTM for Multi-Scale, Multi-Lead Agricultural Drought Prediction

Abstract

The accuracy of deep-learning drought prediction models is normally determined by the ability of the forecasted values to match actual values, which can be a used to evaluate prediction performance. Standardized Precipitation Index (SPI) is the major agriculture drought index and calculated on overlapping rainfall totals, to accumulate extended periods and generate less volatile time series. It is used to predict adequate persistence forecast carrying forward the current value with increasing accumulation period to reflect model predictive ability. In this work,  we forecast SPI for four accumulation periods (3, 6, 9, and 12 months) and for three lead times (1, 3, and 6 months) at 55 grid locations in the Jaisalmer district of Rajasthan, India, employing a Long Short-Term Memory network with temporal attention mechanism. The main model utilized precipitation anomalies, SPI, and seasonal calendar features, while the role of satellite-derived indices is examined in separate sensitivity analysis. . The hyperparameters are tuned using validation samples within each walk-forward sample split. All twelve scale–lead pairs are assessed relative to persistence and climatology using Diebold–Mariano tests in three walk-forward sample splits. For a one-month forecast lead, the absolute accuracy is very high for SPI-3, SPI-6, and SPI-9, but not as high for SPI-12, with the R² values being 0.67, 0.68, 0.65 and 0.33, respectively. With respect to persistence, there are only two configurations that outperformed it in all three walk-forward splits while maintaining the positive absolute R²: SPI-3 at a one-month forecast lead and SPI-6 at a three-month forecast lead. Drought classification followed the same trend. The ROC-AUC is higher than 0.85 in eight configurations, while AW-LSTM outperformed persistence only in four configurations.