BiLSTM-Kalman framework for precipitation downscaling under multiple climate change scenarios

Abstract

Traditional downscaling techniques often fail to accurately represent critical extremes necessary for effective adaptation planning. This paper introduces the first application of Bidirectional Long Short-Term Memory (BiLSTM) networks with an adaptive Kalman filter for multi-scenario, high-resolution precipitation downscaling. We applied our methodology to Tehran, Iran, and systematically compared and ranked the performance of different CMIP6 projections, with the best performing model being MIROC (NSE: 0.902, R2: 0.91, RMSE: 7.76). The optimized BiLSTM network alone demonstrated strong performance (R2: 0.638, KGE: 0.684), with the adaptive Kalman filter dynamically adjusting its parameters according to precipitation intensity. Our novel contributions are a symmetric dependence loss for predicting extremes and graduated correction using percentiles. Examination of the Shared Socioeconomic Pathways (SSPs) 1 to 5 revealed surprising findings: the SSP1-2.6 (more sustainable) pathway predicted the highest extremes, with a 24.3% increase in 99th percentile intensity over the past. SSP2-4.5, SSP3-7.0, and SSP5-8.5 had increases of 17.8%, 16.5%, and 21.1%, respectively. Generated Intensity–Duration–Frequency curves indicated dramatic changes for short-duration events (10–30 min) under SSP5-8.5 with essential implications for infrastructure planning. Extreme precipitation events (> 95th percentile) revealed a frequency increase from 2.1 to 3.5% for SSP1-2.6 for events exceeding 20 mm/day. The integrated framework effectively translates coarse climate model outputs into practical engineering tools, providing the required quantitative information for planning climate-resilient infrastructure.