A recent study in Big Earth Data presents CA-MTransU-Net, an innovative deep learning architecture designed to map forest burned areas (FBAs) efficiently and accurately. Traditional FBA detection using optical imagery is frequently hindered by cloud cover, while conventional transformer-based models suffer from high computational costs. To resolve this, researchers integrated cloud-penetrating Sentinel-1 SAR data with Sentinel-2 optical imagery using a novel dynamic cloud-weighting approach. Furthermore, the model employs a Compact Linear Attention Mechanism (CLAM) alongside a Mixture-of-Experts (MoE) framework, which drastically reduces computational complexity while capturing critical global spatial dependencies. Experimental results demonstrate that CA-MTransU-Net achieved a superior mIoU of 87.00% and faster inference speeds compared to benchmark algorithms like U-Net, SegFormer, and DeepLabv3+. Ultimately, this architecture provides a highly scalable and robust solution for post-fire damage assessment in cloud-prone and heterogeneous landscapes.
Citation
Tahermanesh, S., Jamali, A., Moghimi, A., Mohsenifar, A., Khankeshizadeh, E., & Mohammadzadeh, A. (2026). CA-MTransUNet: Cloud-Aware Mixture-of-Experts Linear Transformer U-Net for forest burned area (FBA) mapping using Sentinel-1 and Sentinel-2 images. Big Earth Data, 1–30. https://doi.org/10.1080/20964471.2025.2598994
Abstract
Wildfires are becoming more frequent and intense, which highlights the need for precise and effective forest burned area (FBA) detection. Current burn mapping approaches are hindered by challenges including the integration of multimodal datasets, high computational complexity of traditional attention mechanisms in segmentation models, and cloud contamination in optical satellite imagery. To address these issues, we proposed the Cloud-Aware Mixture-of-Experts Linear Transformer U-Net (CA-MTransU-Net). Our model integrates Sentinel-1 SAR and Sentinel-2 optical satellite data using a novel dynamic weighting approach, employs a computationally efficient Mixture-of-Experts (MoE) linear attention mechanism to effectively capture global feature dependencies, and incorporates a cloud-weighting method specifically designed to reduce the adverse impacts of cloud cover in optical satellite data. The developed architecture significantly outperformed several well-known segmentation algorithms, including U-Net, ResNet, SegFormer, TransU-Net, PSPNet, and DeepLabv3+, achieving the highest mean Intersection-over-Union (mIoU) of 87.00%, surpassing baseline models by an average of +6.29%. It also demonstrated superior computational efficiency with faster inference speeds (6.26 ms) compared to conventional transformer-based models like SegFormer (7.81 ms) and TransU-Net (13.17 ms). Despite its achievements, the model exhibits higher peak memory usage, which may limit deployment in resource-constrained environments. Additionally, like other tested models, it occasionally misclassified water bodies as burned areas.
#Deep learning; #burned area mapping; #TransU-Net; #remote sensing #multi-sensor satellite imagery
Big Earth Data is an interdisciplinary Open Access journal which aims to provide an efficient and high-quality platform for promoting the sharing, processing and analyses of Earth-related big data, thereby revolutionizing the cognition of the Earth's systems. The journal publishes a wide range of content, including Research Articles, Review Articles, Data Notes, Technical Notes, and Perspectives. It is now included in ESCI (IF=3.8, Q1), Scopus (CiteScore=9.0, Q1), Ei Compendex, GEOBASE, and Inspec. Starting from 2023, Big Earth Data has announced a new award series for authors: Best and Outstanding Paper Awards.
