SpaTempNet: Deep Spatiotemporal Forecasting of River Morphological Evolution of the Padma River in Bangladesh

Abstract

Rivers are dynamic geomorphological systems experiencing continuous transformation through erosion and sediment transport. In Bangladesh, these processes displace 50,000-200,000 people annually, causing severe losses along the Padma River according to Natural Resources Defense Council (NRDC). Traditional monitoring via field surveys and manual remote sensing proves costly, spatially limited, and unsuitable for multi-year pre-diction. This study presents a comprehensive framework, termed SpaTempNet, for forecasting river morphological evolution from freely available satellite data using spatiotemporal deep learning. A 38-year time-series (1987–2025) of binary water masks was constructed from Landsat (5 TM, 7 ETM+, 8 OLI) and Sentinel (Sentinel-2 MSI, Sentinel-1 SAR) imagery via Google Earth Engine. Water extraction used Modified Normalised Difference Water Index (MNDWI). The proposed Bidirectional ConvLSTM gap-filling model achieved IoU = 0.7366, outperforming classical methods. Five spatiotemporal architectures (ConvLSTM, U-Net+LSTM, Attention U-Net+ConvLSTM, Swin Transformer, ViT-based model) were evaluated across yearly, quarterly, and bi-monthly resolutions. Hybrid CNN-LSTM models consistently outperformed pure transformers. Attention U-Net+ConvLSTM achieved best yearly performance (IoU = 0.7005); U-Net+LSTM led bi-monthly prediction (IoU = 0.7791). Statistical analysis quantified mean annual migration of 255.4 m yr−1 with 1998 extreme of 1,485.5 m yr−1. An expansion of about 290 km2 was projected, and a spatially explicit risk map was generated through long-term forecasting (2026-2040). Results demonstrate that freely available satellite imagery with deep learning can provide practical, scalable framework for riverbank hazard monitoring and disaster management.