Black Summer highlighted the vulnerability of coastal wetlands to bushfires. We evaluated the reliability of NSW Fire Extent and Severity Mapping (FESM) - a product calibrated from terrestrial settings – for mangroves and saltmarsh. Using independent reference data of burnt and unburnt wetland extent, we found that, despite high statewide overall accuracy (94–95%), FESM showed very low precision (13–23%) and low F1-scores (0.22–0.32), indicating substantial overestimation of burnt areas in mangrove and saltmarsh ecosystems. To address this, a machine learning framework using Sentinel-2 imagery and DEM-derived variables was developed in the Clyde River estuary. A new Random Forest model achieved an overall accuracy of 82.5% and an F1-score of 0.83, while a Convolutional Neural Network model achieved a comparable accuracy of 82.7% and an F1-score of 0.83. Model robustness was further assessed through external validation across three independent estuaries in the region, demonstrating spatial transferability beyond the training area. The selected model was subsequently applied across coastal wetlands in NSW to generate statewide estimates of burnt and unburnt mangrove and saltmarsh areas. This study demonstrates the importance of validating existing fire severity products and demonstrates new, more robust approaches for mapping and quantifying fire impact in coastal wetlands.
Key Words: Mangrove; Saltmarsh; Machine Learning; Black summer; Bushfire; Natural hazard