An uncertainty-aware Deep Learning approach for rapid burned area mapping using high-resolution PlanetScope multispectral imagery

Abstract

The increasing frequency of wildfires in wetlands demands rapid and accurate monitoring tools to mitigate regional ecological damage and assess carbon emission impacts. While Deep Learning (DL) models offer high precision for environmental monitoring, their operational deployment by environmental agencies is often hindered by the prohibitive cost and time required to generate dense, pixel-wise manual annotations. To address this labeling bottleneck, which critically delays disaster response and restoration policies, this work proposes an uncertainty-aware Weakly Supervised Semantic Segmentation (WSSS) framework tailored for high-resolution multispectral imagery. Our approach relies solely on image-level tags (e.g., “burned” or “unburned”) to train segmentation models, significantly democratizing and accelerating the mapping process. Technically, we adapt the Puzzle-CAM architecture to handle RGB-NIR data and integrate an Imbalanced Class and Uncertainty-aware loss ($L_{ICU}$) during the training stage. This loss function is specifically designed to mitigate the label noise inherent in the WSSS, particularly at the fuzzy boundaries between burned areas and spectrally similar wetland features. Experimental results demonstrate that our method produces sharp, boundary-adherent segmentation masks, achieving an IoU of 83.74% for the burned class. Notably, our approach outperforms recent contrast-based WSSS baselines and excels in detecting small, fragmented fire scars (1%–20% coverage). Furthermore, while a performance gap of approximately 11 percentage points remains when benchmarked against cutting-edge fully supervised foundation models, these findings suggest that uncertainty-aware WSSS is a viable, cost-effective alternative for operational fire monitoring, providing robust spatial data for environmental management, carbon accounting, and data-driven ecosystem analysis.