Assessment and forecasting of drought in Nepal using long short-term memory model

Abstract Accurate drought forecasting is essential for effective water resource management, agricultural planning, and disaster risk reduction, particularly in the context of climate change. Nepal is significantly impacted by drought, yet forecasting efforts in this region remain underexplored. This study aims to analyze drought trends using the Standardized Precipitation Evapotranspiration Index (SPEI) at five timescales (SPEI-1, SPEI-3, SPEI-6, SPEI-9, and SPEI-12) and to forecast these indices using machine learning techniques. Meteorological data from 59 stations across Nepal were used to calculate SPEI values, with data from 1980 to 2013 used for training the LSTM model and data from 2014 to 2023 for testing. Meteorological stations were grouped into seven homogeneous climatic regions using K-means clustering. We employed a long short-term memory (LSTM) model, well-suited for capturing temporal patterns, to predict drought conditions. The model’s performance was evaluated using metrics such as the coefficient of determination (R 2 ), root mean square error (RMSE), and mean absolute error (MAE). Our findings reveal a significant decreasing trend (approximately 50%) in SPEI values at shorter timescales (e.g., SPEI-1), with the trend becoming less pronounced (30–37%) as the timescale increases to SPEI-3, SPEI-6, SPEI-9, and SPEI-12. R² values across all stations ranged from approximately 0.25–0.75 for SPEI-3, 0.45–0.82 for SPEI-6, 0.60–0.88 for SPEI-9, and 0.70–0.90 for SPEI-12. Forecasting results demonstrate improved model performance at longer timescales, with SPEI-12 predictions showing the highest accuracy and SPEI-3 the lowest. These results show the potential for using LSTM models in drought forecasting to support water resource management, enhance agricultural resilience, and mitigate the socio-economic and environmental impacts of water scarcity.