Satellite-Based Forest Fire Mapping Using Group and Shuffle Network Embeddings

Forest fire mapping plays a vital role in environmental monitoring and disaster response. While satellite imagery and deep learning have advanced automated fire detection, many existing approaches remain limited in accurately segmenting small-scale and irregular fire patches. The purpose of this study is to develop and evaluate GSDABNet, a novel deep semantic segmentation architecture for improved delineation of wildfire regions using Landsat-8 imagery. The model integrates depthwise asymmetric bottleneck (DAB) modules with a group-and-shuffle mechanism, enhancing feature diversity and spatial–spectral representation while reducing redundant global patterns. The methodological workflow includes preprocessing Landsat-8 imagery by selecting SWIR-1, SWIR-2, and Blue bands, building the proposed encoder–decoder framework with GS and DAB modules, and training on a publicly available global wildfire dataset. Model evaluation was performed using accuracy, precision, recall, F1-score, and mean Intersection over Union (mIoU), with additional ablation studies on GS placement, group size, and shuffle permutations to assess architectural contributions. Results show that the optimal configuration two groups with a (0,1,2,4,3) shuffle applied after the fifth DAB block achieved an mIoU of 0.6476, accuracy of 0.9957, and F1-score of 0.7201, outperforming multiple state-of-the-art benchmark models. These findings confirm GSDABNet as a highly accurate and computationally efficient framework for wildfire segmentation, contributing to improved early detection, effective risk mitigation, and sustainable forest management.