Project aim

Forecast where the shoreline will be three months ahead, with uncertainty, so erosion hotspots can be watched before infrastructure is lost.

  • <8 mForecast MAE
  • 3 monthsForecast horizon
  • 4Imagery sources
  • DSASValidation standard
Illustrative map for Coastal Erosion Digital Twin Illustrative graphic

Overview

The digital twin combines Sentinel-1/2, Landsat, UAV orthomosaics and wave climatology. A Swin-UNETR backbone extracts shoreline features and a Bayesian LSTM forecasts their position, reaching MAE below 8 m for 3-month forecasts when validated against DSAS transects.

The challenge

Shorelines move with tides, seasons and storms, so a single image can mislead. Different sensors see the coast at different resolutions and times, and a forecast is only useful if it reports how uncertain it is.

Data & inputs

  • Sentinel-1/2
  • Landsat
  • UAV orthomosaics
  • Wave climatology
  • DSAS transects

Approach

  1. 01

    Multi-sensor archive

    Sentinel-1/2, Landsat and UAV orthomosaics provide shoreline observations at different scales.

  2. 02

    Wave climatology

    Wave conditions are added as drivers of change.

  3. 03

    Swin-UNETR backbone

    A transformer-based segmentation backbone extracts the shoreline from imagery.

  4. 04

    Bayesian LSTM

    A Bayesian LSTM forecasts shoreline position with uncertainty.

  5. 05

    DSAS validation

    Forecasts are checked against Digital Shoreline Analysis System transects.

Results

3-month shoreline forecasts reach a mean absolute error below 8 m against DSAS transects.

Validation

Validated against DSAS (Digital Shoreline Analysis System) transects.