Adaptive Weighted LSTM For Long-Lead Drought Forecasting Using Hydroclimatic Time Series In Arid Regions

Abstract

This study was conducted in Jaisalmer district, located in the western part of Rajasthan, India, an arid region highly vulnerable to persistent drought conditions. The area experiences low and erratic rainfall, making it critical for early drought detection and long-term hydrological forecasting. This study proposes an Adaptive Weighted Long Short-Term Memory (AW-LSTM) model to forecast the Standardized Precipitation Index (SPI) using high-resolution gridded hydroclimatic variables such as precipitation, vegetation (VCI), and temperature (TCI). AW-LSTM integrates dynamic attention weights that prioritize relevant time steps and input features to
enhance the long-term drought forecasting accuracy. The model was trained using data from 1991 to 2020 and validated from 2021 to 2023 across lead times of 1 to 12 months. The AW-LSTM model achieved a high level of predictive accuracy, with an R² of 0.98, RMSE of 0.02, and Mean Absolute Error (MAE)of 0.08, successfully capturing the beginning, duration, and severity of droughts. The model revealed that the western and southwestern parts of Jaisalmer were the most drought-prone, whereas the northeastern and central-eastern regions received relatively high rainfall. These findings support the development of real-time drought monitoring systems and enhance adaptive decision making for water resource planning in arid regions.