Weakly Supervised Fine-Resolution Wildfire Fuel Mapping From Spatially Coarse Labels and Multisource Remote Sensing Data

Abstract

Accurate large-scale wildfire fuel mapping is essential for fire-risk assessment and ecosystem management, yet fine-resolution training labels remain costly and scarce. This study presents L2HFuelNet, a weakly supervised resolution-enhancement framework that uses spatially inexact 30 m labels to generate 10 m vegetation-based fuel-type predictions in seven aggregated categories. Sentinel-2 multispectral imagery, Sentinel-1 synthetic aperture radar backscatter, and topographic variables are integrated through a dual-encoder architecture combining a Swin Transformer with an Atrous Spatial Pyramid Pooling-based convolutional encoder. A composite low-to-high loss comprising weighted cross-entropy, intersection-over-union, and confidence–vague distance terms is employed to address class imbalance and supervisory ambiguity caused by the resolution mismatch. Province-wide evaluation across the heterogeneous landscapes of Alberta, Canada, yielded an overall accuracy of 0.91 and a mean intersection-over-union of 0.84, outperforming the evaluated baselines and improving discrimination among spectrally similar classes, including Mixedwood, Pine, and Aspen. Coarse-supervision analysis further characterized class-specific label–prediction discrepancies and their influence on mapping performance. These findings demonstrate the feasibility of province-scale fine-resolution fuel characterization using coarse supervisory labels and multisource remote sensing data. The resulting categories are designed to support landscape-scale fuel assessment and can complement detailed fire behavior prediction (FBP) information when subclass-specific fire-behavior parameters are required.