CA-MTransUNet: Cloud-Aware Mixture-of-Experts Linear Transformer U-Net for forest burned area (FBA) mapping using Sentinel-1 and Sentinel-2 images

Sahand Tahermanesh, Ali Jamali, Armin Moghimi, Amin Mohsenifar, Ehsan Khankeshizadeh, and Ali Mohammadzadeh

Citation

Please kindly cite the paper if this code is useful and helpful for your research.

  @article{Tahermanesh2026,
    title = {CA-MTransUNet: Cloud-Aware Mixture-of-Experts Linear Transformer U-Net for forest burned area (FBA) mapping using Sentinel-1 and Sentinel-2 images},
    author = {Tahermanesh, S., Jamali, A., Moghimi, A., Mohsenifar, A., Khankeshizadeh, E., & Mohammadzadeh, A. },
    journal = {Big Earth Data},
    volume = {},
    pages = {1–30},
    year = {2026},
    issn = {},
    doi = {https://doi.org/10.1016/j.jag.2023.103332},
    url = {https://www.tandfonline.com/doi/full/10.1080/20964471.2025.2598994?src=}
  }

Figure 1. (a) Overview of the proposed CA-MTransU-Net architecture, illustrating hierarchical feature encoding, MoE-based global attention, and decoding with addition-based skip connections; (b) CLAM, illustrating the efficient calculation of global attention via kernel-based approximation; and (c) the MoE transformer module consisting of token-mixing through CLAM and channel-mixing using a linear layer (i.e., MLP).

Figure 2. Overall architecture of (a) Conventional attention mechanisms, (b) the proposed CLAM adapted from Performer (Choromanski et al., Citation2020). The proposed attention module employs a kernel-based approximation (ReLU) to efficiently compute attention. It reorders computations by first aggregating the values 𝑉 using the keys (𝐾⁢ ′.𝑉) and then weighting this aggregation with the queries (𝑄⁢ ′⁢(𝐾⁢ ′.𝑉)), and (c) a weighted combination of several CLAM-based attention modules.

Figure 3. An example of the estimated cloud probability map; (a) Sentinel-2 image, and (b) cloud probability.

Figure 4. Visualization of expert attention maps generated by the CLAM attention mechanism embedded within the MoE transformer for several representative samples. (a) RGB post-fire Sentinel-2 images, (b) ground truth maps where burned areas are shown in white, (c) attention map of the first CLAM expert, (d) attention map of the second clam expert, and (e) attention map of the third CLAM expert.

License

Copyright (c) 2026 Ali Jamali. Released under the MIT License. See LICENSE for details.