Predicting riverbank erosion and accretion using multi-level LSTM-based deep models and satellite-derived land cover indices

Abstract

This research presents a multi-level LSTM-based deep learning framework, centered on a Bidirectional LSTM–GRU hybrid with attention, which predicts riverbank erosion and accretion by a long-term satellite-based dataset spanning from 1988 to 2025. Data covering seven rivers and multiple zones were acquired from Landsat 5, Landsat 7, and Sentinel-2 satellites. The research applied advanced geospatial processing to generate annual water masks, riverbank polygons, and extract crucial land cover indices such as NDWI, NDVI, NDBI, SAVI, MSI, NBR, and BAI, which are complemented by elevation, slope, and soil texture data to characterize riverbank environments using Google Earth Engine. Furthermore, the 1988 imagery served as a baseline for quantifying erosion and accretion changes annually on left and right banks. Data processing included cloud masking, vectorization of water bodies, best line detection for riverbank delineation, and spatial classification of banks. For each river zone and bank side, the erosion and accretion areas were spatially and temporally quantified. Moreover, a hierarchical sequence architecture is used to model temporal dependencies on a global, river, zone, and bank level using Bidirectional LSTM-GRU with attention layers. To ensure robust temporal continuity for model training, data were preprocessed through category quantization based on global tertiles and sequence interpolation to fill the missing years.The model predicts the erosion and accretion dynamics, and the effectiveness of the model is evaluated against baseline models (standard LSTM, Random Forest, XGBoost, LightGBM) and tested with the help of detailed trend analysis over time, heatmaps of the correlation, and confusion matrices. Additionally, multi-scale spatial and temporal variability and associations among rivers and bank zones were revealed by visualizations. The framework integrates remote sensing, geospatial analytics, and advanced deep learning to enable scalable and interpretable riverbank change prediction which provides valuable insights for river management, conservation, and risk mitigation strategies.