Neural Networks Learn to Sketch Tomorrow’s Coastline From Decades of Landsat Views

A hybrid deep-learning model has been used to forecast the water–land configurations of a Brazilian coastline from just seven satellite snapshots spread over three decades, in a proof-of-concept study that tests how far sparse satellite archives can stretch when the coastal monitoring record is thin. The work, carried out on the roughly 30-kilometre shoreline of Icapuí in the state of Ceará, northeastern Brazil, combined multi-temporal Landsat imagery, an automated water-index extraction procedure, and a hybrid convolutional neural network–long short-term memory (CNN–LSTM) architecture to predict the next coastal state in a temporal sequence.

The study, published in Earth Science Informatics, tackles a persistent gap in coastal forecasting. Process-based numerical models such as Delft3D and XBeach can simulate hydrodynamics and sediment transport in detail, but they demand extensive calibration data, detailed bathymetry and environmental forcing variables that many data-scarce regions simply lack. Classical statistical tools, including end point rate, net shoreline movement and linear regression rate, are interpretable and widely used, yet they struggle with the nonlinear, spatially heterogeneous behaviour that real coastlines display. The researchers set out to explore whether a hybrid deep-learning framework could learn both spatial patterns and temporal dynamics directly from satellite-derived data, even when only a handful of acquisition dates were available.

The pipeline begins with the Modified Normalized Difference Water Index, or MNDWI, a spectral index calculated from the green and shortwave-infrared bands of Landsat 5 TM and Landsat 8 OLI imagery. MNDWI enhances the contrast between open water and land surfaces while suppressing noise from built-up areas, which makes it a standard choice for automated shoreline delineation. After radiometric, atmospheric and geometric preprocessing co-registered to an orthorectified 2009 reference image, each MNDWI raster was passed through a Gaussian filter and threshold-based binarization to produce a binary water–land mask, from which a vector shoreline could be derived for spatial interpretation.

Two separate Landsat datasets underpinned the two modelling workflows. A linear-regression experiment used six observations acquired at roughly six-year intervals between 1985 and 2015, extrapolating first-order trends and generating a prognostic shoreline scenario for 2021 from the 2009 and 2015 images. The CNN–LSTM experiment used an independently constructed sequence of seven observed scenes acquired between 1988 and 2018, at approximately five-year intervals and concentrated in August and September. Crucially, the authors emphasize that no regression-generated or synthetic imagery entered the deep-learning training or evaluation; the two components are treated as complementary rather than as a controlled head-to-head comparison, since they used different temporal series, prediction targets and evaluation domains.

The hybrid architecture chains two complementary network types. Four convolutional blocks, containing 64, 128, 256 and 512 filters with 3-by-3 kernels, ReLU activation, 2-by-2 max-pooling and dropout of 0.20, learn latent spatial representations from each MNDWI raster. These learned image-domain features are not individually labelled with physical meanings such as beach width or erosion rate; they are encodings of local spatial contrasts and arrangements around the water–land interface. The flattened feature maps are then passed to three stacked LSTM layers, each with 1024 memory cells, which model the sequential evolution of these representations across the ordered raster series. A final dense layer with 512 neurons reconstructs the MNDWI raster corresponding to the next time step, and the shoreline is derived from that predicted raster only afterwards through the same post-processing workflow.

To test the value of temporal context, the team ran two experimental scenarios: a two-to-one configuration, in which two consecutive observed rasters predict the third, and a three-to-one configuration using three inputs to predict the fourth. Because the study area spans geomorphologically heterogeneous terrain—sandy beaches, active and inactive cliffs, dune fields, lagoons and estuarine environments—the coast was divided into five sectors (East, Central-East, Central-West A, Central-West B and West), each processed as a complete single-band raster without patch-based tiling. Model training ran for 150 epochs, after which prediction errors stabilized across sectors.

Performance was evaluated with mean absolute error and root mean square error in normalized MNDWI raster-value units, comparing predicted rasters against the observed Landsat-derived targets for 2003, 2008, 2013 and 2018. Across the five sectors, MAE ranged from 0.044 to 0.066 in the two-timestep configuration and from 0.039 to 0.065 in the three-timestep configuration, with RMSE values from 0.054 to 0.088 and 0.051 to 0.086 respectively. The Central-East sector recorded the highest errors, consistent with its greater geomorphological complexity. In most sectors the three-timestep setup produced lower reconstruction errors, suggesting that the extra preceding observation supplied useful temporal context for raster reconstruction within the analysed sequence—though the authors caution that this pattern should not be read as general evidence of strong long-term temporal dependence.

The team is careful to keep these metrics in their proper domain. Because the errors measure differences in normalized MNDWI intensity, they quantify image reconstruction accuracy and cannot be converted into shoreline-position errors in metres; MNDWI intensity differences have no fixed relationship with physical displacement, and a positional assessment would require vectorizing both predicted and observed boundaries and comparing them geometrically, a validation step not performed in the original experiment. The 30-metre nominal resolution of Landsat imagery likewise cannot serve as a conversion factor. Several extracted shorelines showed apparent changes whose magnitude approaches the pixel size itself, and the authors also flag tidal stage, wet sand, wave run-up, mixed pixels and sensor differences as sources of positional uncertainty, noting that no explicit tidal correction was applied.

The most forward-looking output is an out-of-series forecast for 2023, generated five years beyond the last observed acquisition. With no corresponding observed target available, this configuration was treated as an exploratory model-derived scenario. Its behaviour varied spatially: the East sector showed localized divergence from the historical range of shoreline configurations, while the Central-East and Central-West sectors stayed broadly within the pattern of the observed series. The authors stress that none of these divergences or agreements can be independently verified as real morphological change or forecasting accuracy; they may reflect model behaviour under extrapolation, extraction uncertainty or acquisition-dependent environmental conditions.

The study’s value lies in its methodological honesty as much as its results. The CNN–LSTM component was built on only seven independent temporal observations, and the subdivision into sectors did not increase that count; the overlapping chronological input–target windows—five possible for the two-timestep setup and four for the three-timestep one—cannot be treated as fully independent training samples, and no separate temporal holdout set was maintained. Some implementation details, including the optimizer, learning rate and source code, were not preserved in the archived records, preventing full reproducibility. With those caveats, the work demonstrates that a hybrid spatial–temporal deep-learning framework can operate on satellite-derived spectral-index rasters under sparse temporal conditions, and it charts what a credible next generation of coastal forecasting systems will need: denser observations, independent temporal validation, metric-domain shoreline evaluation, physical forcing variables such as waves, tides and winds, and controlled comparisons against alternative statistical and machine-learning approaches. Until then, the model-derived shorelines of Icapuí stand as possible trajectories rather than confirmed predictions of where the coast is heading.

Subject of Research: CNN–LSTM-based spatio-temporal forecasting of MNDWI-derived coastal states from sparse multi-temporal Landsat imagery in Icapuí, northeastern Brazil

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

Article References: de Sousa, W. R. N., Souto, M. V. S., Salgueiro, A. R. G. N. L., & de Castro, A. F. (2026). CNN–LSTM-Based Spatio-Temporal Forecasting of MNDWI-Derived Coastal States: A Multi-Temporal Landsat Study in Northeastern Brazil. Earth Science Informatics, 19(12), Article 227. https://doi.org/10.1007/s12145-026-02283-8

Image Credits: AI Generated

DOI: 10.1007/s12145-026-02283-8

Keywords: CNN-LSTM, shoreline forecasting, Landsat, MNDWI, remote sensing, northeastern Brazil, coastal monitoring, deep learning, spatio-temporal modelling, Icapuí, Ceará, proof of concept