# Swamp-AI: a deep learning model for monitoring wetlands change across the globe

## TL;DR

Swamp-AI is a deep learning model trained to monitor global wetland changes using remote sensing data, offering high accuracy for inaccessible regions.

## Contribution

Swamp-AI introduces a globally trained deep learning model with a novel annotated database for wetland monitoring.

## Key findings

- Swamp-AI achieved 93.7% overall accuracy in identifying wetland changes.
- The model was trained on a diverse dataset covering various wetland types and seasonal imagery.
- Results suggest Swamp-AI is a generalizable tool for global wetland monitoring.

## Abstract

Cost-effective and rapid identification of changes to wetland extent is increasingly critical given their decline in coverage worldwide over the past several decades. In the case of highly remote or otherwise physically inaccessible wetlands, remote sensing often proves the only feasible method for long-term monitoring. Recently, remote sensing approaches incorporating machine learning and deep learning (DL) have gained prominence as tools for monitoring changes to wetland extent. Here, we present “Swamp-AI” a DL model trained on wetland locations from all over the world. We devised a unique annotation system leveraging multiple global datasets to create an annotated imagery database of wetlands drawn from across the globe. The scenes selected for annotation were carefully chosen to encompass a wide variety of wetland types, including coastal and inland systems. To account for seasonality, imagery acquired throughout the year was included. This annotated database was used to train 15 candidate DL models. The highest performing model, referred to as “Swamp-AI”, achieved averaged scores of 93.7% overall accuracy, 79.4% producer’s accuracy, 93.2% user’s accuracy, and 74.6% intersection over union across the test sites. These results suggest that Swamp-AI represents a promising generalizable tool for monitoring changes in the extent of wetlands globally.

The online version contains supplementary material available at 10.1038/s41598-026-39257-1.