Scholarlink

Deforestation significantly contributes to climate change, driving efforts to analyze its causes and create accurate segmentation maps. Recent advances show that Vision Transformers outperform traditional Convolutional Neural Networks in many vision tasks. However, their use in remote sensing remains limited due to heavy computational needs and dependence on large labeled datasets. This work proposes an attention-driven architecture, TransU-Net++, designed for semantic segmentation to map deforestation in the Atlantic Forest and Amazon Rainforest regions. The framework combines Heterogeneous Kernel Convolution, U-Net, attention gates, and ViT to enhance feature learning and segmentation precision. The model is compared with leading networks—U-Net [9], Attention U-Net [15], Swin U-Net [8], SegNet [17], ICNet [18], ENet [14], R2Unet [19], Attention U-Net-2 [20], and TransU-Net [10]. On the 4-band Atlantic Forest dataset, TransU-Net++ achieves an accuracy of 87.51%, precision of 72.36%, recall of 89.87%, and an F1-score of 82.10%, outperforming models such as ICNet (65.49%), ENet (67.87%), SegNet (69.01%), U-Net (80.96%), U-Net+++ (80.23%), R2U-Net (81.71%), Attention U-Net (82.16%), TransU-Net (81.40%), Swin U-Net (81.93%) and ResUnetA [16].

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

TransU-Net++: Attention-Gated Transformer U-Net for Deforestation Mapping in South American Forests

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Akshay Poojary,
  • Varsha Sajjanavar,
  • N. Abhishek,
  • Indira Bidari,
  • Sathyanand Chickerur

摘要

Deforestation significantly contributes to climate change, driving efforts to analyze its causes and create accurate segmentation maps. Recent advances show that Vision Transformers outperform traditional Convolutional Neural Networks in many vision tasks. However, their use in remote sensing remains limited due to heavy computational needs and dependence on large labeled datasets. This work proposes an attention-driven architecture, TransU-Net++, designed for semantic segmentation to map deforestation in the Atlantic Forest and Amazon Rainforest regions. The framework combines Heterogeneous Kernel Convolution, U-Net, attention gates, and ViT to enhance feature learning and segmentation precision. The model is compared with leading networks—U-Net [9], Attention U-Net [15], Swin U-Net [8], SegNet [17], ICNet [18], ENet [14], R2Unet [19], Attention U-Net-2 [20], and TransU-Net [10]. On the 4-band Atlantic Forest dataset, TransU-Net++ achieves an accuracy of 87.51%, precision of 72.36%, recall of 89.87%, and an F1-score of 82.10%, outperforming models such as ICNet (65.49%), ENet (67.87%), SegNet (69.01%), U-Net (80.96%), U-Net+++ (80.23%), R2U-Net (81.71%), Attention U-Net (82.16%), TransU-Net (81.40%), Swin U-Net (81.93%) and ResUnetA [16].