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

Area of Science:

• Climate Science
• Hydrology
• Artificial Intelligence

Background:

• Traditional downscaling methods struggle to capture extreme precipitation events crucial for adaptation planning.
• Accurate high-resolution precipitation data is essential for understanding climate change impacts and infrastructure resilience.

Purpose of the Study:

• To introduce and validate a novel downscaling framework using Bidirectional Long Short-Term Memory (BiLSTM) networks and an adaptive Kalman filter.
• To assess and rank different Coupled Model Intercomparison Project Phase 6 (CMIP6) projections for precipitation downscaling.
• To analyze future extreme precipitation changes under various Shared Socioeconomic Pathways (SSPs) and their implications for infrastructure.

Main Methods:

• Application of Bidirectional Long Short-Term Memory (BiLSTM) networks coupled with an adaptive Kalman filter for precipitation downscaling.
• Systematic comparison and ranking of CMIP6 climate model projections using performance metrics (NSE, R², RMSE).
• Development of a symmetric dependence loss function and graduated correction using percentiles for predicting extreme events.

Main Results:

• The MIROC CMIP6 model exhibited the best performance for downscaling in Tehran (NSE: 0.902, R²: 0.91).
• The optimized BiLSTM network achieved strong performance (R²: 0.638, KGE: 0.684), with the Kalman filter adapting to precipitation intensity.
• Contrary to expectations, the sustainable SSP1-2.6 pathway projected the highest increase in extreme precipitation intensity (24.3% for the 99th percentile).
• Intensity-Duration-Frequency curves showed significant changes for short-duration events under SSP5-8.5, impacting infrastructure planning.
• Extreme precipitation events (>95th percentile) are projected to increase in frequency across all SSPs.

Conclusions:

• The integrated BiLSTM-Kalman filter framework provides a robust tool for high-resolution precipitation downscaling and extreme event prediction.
• The findings highlight the critical need for infrastructure adaptation, even under sustainable development pathways.
• This methodology effectively translates coarse climate model outputs into actionable data for climate-resilient infrastructure development.