Move through a changing shoreline.
A map and time control make the record tangible, allowing a researcher or planner to move from observed history into possible futures without confusing one for the other.
A Softmachines work · Coastal intelligence
The coast is a living record.
REDON makes shoreline movement easier to see, question, and act on, without making uncertainty disappear.
The premise
A forecast is not a verdict.
Coastlines change through tide, storm, sediment, rivers, vegetation, and human decisions. A single confident-looking prediction would hide more than it explains.
REDON brings satellite-derived observations, transects, alternative models, evaluation, and uncertainty into one place, turning a distant risk into evidence people can examine together.
A system for attention
History, futures, and the questions between them.
Keep uncertainty usable.
The evaluation view places errors, withheld observations, and transect behavior beside the forecast. It shows where a model matched reality and where it did not, so uncertainty becomes part of the decision rather than a footnote.
Compare, rather than defer.
The analysis view exposes temporal structure and model behavior instead of reducing everything to a single line. Learning-based, trend, and hybrid physical approaches become alternatives a researcher can test, not authorities to obey.
Make time visible on the map.
The forecast timeline lets people watch observed and projected shorelines move through the same geography. Land loss, gain, transects, and confidence remain visible together, connecting statistical change to an actual coast.
Forecasting stack
No single model gets the last word.
REDON trains and compares several statistical, neural, foundation-model, physical, and ensemble approaches. Each model exposes a different assumption about how a shoreline changes, giving researchers alternatives to test against observations rather than one opaque answer.
- Trend-seasonal
- Data-driven baseline
- Separates long-term shoreline movement from recurring seasonal behavior. It provides a transparent reference forecast and makes it easier to see whether a more complex model is actually learning anything useful.
- Linear
- Statistical baseline
- Projects the observed direction of change with a simple linear relationship. Its restraint is valuable: it establishes the minimum performance a more sophisticated approach must exceed.
- Equilibrium
- Hybrid physical
- Encodes an equilibrium-style shoreline response so the forecast is informed by coastal behavior as well as the measured series. It offers a physically interpretable counterpoint to purely learned patterns.
- DLinear
- Neural time series
- Decomposes the sequence into trend and residual seasonal structure, then forecasts them through learned linear layers. It is lightweight enough to remain a useful bridge between statistical baselines and deeper recurrent models.
- LSTM
- Recurrent network
- Learns dependencies across ordered shoreline observations, retaining information from earlier states while processing later ones. It is suited to temporal relationships that a straight trend cannot represent.
- TimesFM
- Foundation model
- Applies a pretrained time-series forecasting model and adapts it to the record of a particular site. This transfers broader temporal knowledge while still requiring local fine-tuning and validation against the coast being studied.
- Ensemble
- Combined forecast
- Brings multiple model families together so the result is less dependent on one set of assumptions. Agreement and disagreement between members become additional evidence about forecast stability.
Training is only the beginning.
REDON organizes shoreline distance through spatial transects, removes flagged outliers during quality assurance, and learns from observations that may be unevenly distributed through time. Multi-transect training allows the system to recognize structure shared across nearby parts of a coast while preserving the behavior of individual transects.
Every run is evaluated against a withheld period the model did not see during training. Blind-validation RMSE, per-transect error, predicted-versus-observed trajectories, and lead-dependent uncertainty show where performance holds and where it degrades.
The selected model then produces a continuous shoreline trajectory and an uncertainty band for animation on the map. Researchers can compare runs, inspect the source data, and reject a technically impressive forecast when its evidence is weak or its assumptions do not fit the site.
AI becomes a partner in attention.
Shoreline change is distributed across years, locations, and uneven observations. REDON uses learning-based forecasting alongside simpler trends and physically informed approaches to surface temporal relationships that are difficult to hold in mind one measurement at a time.
No model is allowed to turn weak evidence into certainty. Researchers remain responsible for the source data, site selection, model choice, validation period, and interpretation of error. The interface keeps those decisions close to the forecast so technical sophistication does not conceal its assumptions.
The larger aim is a shared habit of noticing change before it becomes irreversible: clearer evidence for researchers, planners, and coastal communities, and more time for public choices while choices still exist.
Earlier recognition creates more time to choose well.
Return to works