Voruarn/HCM

ADVEI 2025, Controllable Diffusion Generated Dataset and Hybrid CNN-Mamba Network for Burned Area Segmentation # HCM, ADVEI 2025 - Paper: Controllable diffusion generated dataset and hybrid CNN–Mamba network for burned area segmentation ## Abstract Burned area segmentation (BAS) in remote sensing images (RSIs), as a key task in forest fire monitoring, is crucial for post-disaster damage assessment and forest recovery guidance. However, BAS sample labeling is time-consuming and laborious as it is densely labeled and requires expert knowledge. To address this, we present Mask-Guided Diffusion Transformer (MG-DiT), a controllable BAS sample generation method using given mask to generate diverse and realistic synthetic images, which uses only mask-image pairs during training. We use MG-DiT to generate Synth-CBAS and Synth-Bushfire, the first synthetic BAS datasets based on Landsat8 and Sentinel2 images. The two generated datasets can work as pretraining data for BAS, thus significantly enhancing the performance of various segmentation models, as demonstrated in the experiments. Besides, we also propose a hybrid CNN–Mamba (HCM) network for BAS, which is based on encoder–decoder structure. In HCM’s encoder, we integrate ConvNextV2 blocks (CNN branch) for spatial feature extraction and visual state space (VSS) blocks (Mamba branch) for long-term dependencies modeling. Later, we propose cross-level feature fusion decoder (CFFD) with gated fusion units to effectively fuse multi-scale features from CNN and Mamba branch, thus generating segmentation maps. Comprehensive experiments on two benchmark datasets substantiate that our approach is more effective than the existing cutting-edge methods. Code is available at: github.com. ## Dataset - **CBAS**: - **Bushfire**: ## Related Works Dual backbone interaction network for burned area segmentation in optical remote sensing images , IEEE GRSL 2024. Burned area segmentation in optical remote sensing images driven by U-shaped multistage masked autoencoder , IEEE JSTARS 2024. A novel salient object detection network …