CNN–LSTM-Based Spatio-Temporal Forecasting of MNDWI-Derived Coastal States: A Multi-Temporal Landsat Study in Northeastern Brazil

Accurate forecasting of shoreline evolution is important for coastal monitoring and management, particularly in dynamic and data-scarce environments. This study investigates a modular framework, applied to Icapuí, northeastern Brazil, that integrates multi-temporal Landsat imagery, remote-sensing-based shoreline extraction, linear regression, and a hybrid convolutional neural network–long short-term memory (CNN–LSTM) architecture. Two distinct Landsat temporal datasets were used for the statistical and deep-learning experiments, and no regression-generated imagery was incorporated into CNN–LSTM training or evaluation. Water–land boundaries were derived from observed Landsat imagery using the Modified Normalized Difference Water Index (MNDWI) followed by binary classification. For spatio-temporal modelling, the CNN–LSTM operated directly on sequences of observed MNDWI rasters to predict the raster corresponding to the subsequent time step. Model performance was evaluated using MAE and RMSE in normalized MNDWI raster-value units, which quantify raster reconstruction error rather than shoreline-position error in physical units. Predicted MNDWI rasters were subsequently converted into binary water–land representations from which vector shorelines were derived for spatial interpretation. Across the analysed coastal sectors, some three-timestep configurations produced lower reconstruction errors than their two-timestep counterparts, suggesting that additional temporal context may provide useful information for raster reconstruction within the analysed sequence. The out-of-series 2023 forecast showed spatially variable behaviour among coastal sectors and was treated as an exploratory model-derived scenario because no corresponding observed target was available for independent validation. Overall, the study provides a proof-of-concept assessment of integrating satellite-derived coastal observations with spatio-temporal deep learning under sparse temporal data availability, while highlighting the need for denser observations, independent temporal validation, and metric-domain shoreline evaluation before robust long-term forecasting capability can be established.