Mamba–LSTM for Long-Horizon Shoreline Prediction Using Satellite Imagery

Abstract

Coastal erosion forecasting is central to climate adaptation and risk management. However, shoreline dynamics exhibit complex spatio-temporal patterns with both local variations and long-range dependencies that conventional recurrent models struggle to capture efficiently. We propose a hybrid Mamba-LSTM framework for multi-step coastline forecasting that combines the efficient long-range modeling of state space models with the sequential learning capacity of recurrent networks. Our Mamba-based temporal encoder leverages selective state spaces to capture extended temporal dependencies, followed by LSTM layers for refined sequential aggregation. To further verify predictive performance of Mamba-LSTM, we design a conditional diffusion extension with cross-attention conditioning as the benchmark. Experiments on a shoreline time series dataset demonstrate that our Mamba-LSTM architecture achieves competitive forecasting accuracy compared to pure LSTM and diffusion baselines, while offering improved modeling of longhorizon temporal patterns. For reproducibility, our training configurations and code are available here 11https://github.com/UofgCoastline/ICMLT-2026-Mamba-LSTM-Coastline-Prediction