Scholarlink

The rapid expansion of solar energy adoption has created a growing need for accurate and scalable methods to identify and quantify photovoltaic (PV) installations from satellite imagery, which is critical for informed energy planning and resource allocation. Conventional deep learning architectures, including U-Net and Mask Region-based Convolutional Neural Network (Mask R-CNN), exhibit significant limitations in solar panel segmentation tasks due to class imbalance, insufficient spatial detail preservation, and restricted generalization across varying terrains and environmental conditions. The proposed hybrid model integrates the advanced contextual feature extraction capability of SegFormer with the spatial refinement efficiency of U-Net, resulting in a robust and detail-oriented segmentation framework. To further enhance segmentation quality, the incorporation of Focal Loss and Dice Loss provides improved sensitivity to minority classes and sharper boundary delineation. In addition to accurate segmentation, the model enables precise area estimation of PV arrays, which can be directly leveraged for solar energy potential assessment and yield prediction, supporting data-driven strategies for renewable energy deployment. The proposed model achieves an accuracy of 97.74%, a precision of 92.0%, and a Dice score of 0.8572, outperforming existing segmentation approaches.

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

A Hybrid Transformer-CNN Architecture Integrating SegFormer and U-Net for Enhanced Image Segmentation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • T. Satyanarayana Murthy,
  • Puppala Mangala Tulasi,
  • Bandam Deekshitha,
  • Paduru Akshaya Reddy,
  • Krushi Sirimalle,
  • K. Gangadhara Rao

摘要

The rapid expansion of solar energy adoption has created a growing need for accurate and scalable methods to identify and quantify photovoltaic (PV) installations from satellite imagery, which is critical for informed energy planning and resource allocation. Conventional deep learning architectures, including U-Net and Mask Region-based Convolutional Neural Network (Mask R-CNN), exhibit significant limitations in solar panel segmentation tasks due to class imbalance, insufficient spatial detail preservation, and restricted generalization across varying terrains and environmental conditions. The proposed hybrid model integrates the advanced contextual feature extraction capability of SegFormer with the spatial refinement efficiency of U-Net, resulting in a robust and detail-oriented segmentation framework. To further enhance segmentation quality, the incorporation of Focal Loss and Dice Loss provides improved sensitivity to minority classes and sharper boundary delineation. In addition to accurate segmentation, the model enables precise area estimation of PV arrays, which can be directly leveraged for solar energy potential assessment and yield prediction, supporting data-driven strategies for renewable energy deployment. The proposed model achieves an accuracy of 97.74%, a precision of 92.0%, and a Dice score of 0.8572, outperforming existing segmentation approaches.