Scholarlink

<p>The study proposes a novel semi-supervised deep learning framework called Residual Cross-Fusion Network (ResCrossNet) for hyperspectral deforestation change detection. The model processes a pair of hyperspectral images captured at different time intervals over the same geographical area, using identical spatial-spectral feature extraction pipelines. The Spatial-Spectral Feature Extraction Module (SSFM) incorporates residual spatial and residual channel attention blocks to effectively capture meaningful spatial and spectral information. These extracted features are then subjected to the Cross-Fuse Module (CFM), which increases complementary features and reduces redundancy between the bi-temporal inputs. The fused features are integrated using differencing and concatenation techniques to detect both obvious and subtle changes in the forest landscape. Finally, semantic segmentation is performed using the Refined Object-Contextual Representation (ROCR) module, which models object-level contextual dependencies to increase the accuracy of pixel-level change. Its excellent categorization accuracy of 99.6% when implemented with Python showed its potential for accurate and scalable monitoring of forest change.</p>

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

Semi-supervised residual cross-fusion network for deforestation change detection on hyperspectral imagery with semantic segmentation

  • 期刊论文

  • 发表日期：

    2025年12月20日

查看全文
  • Indira Bidari,
  • Satyadhyan Chickerur

摘要

The study proposes a novel semi-supervised deep learning framework called Residual Cross-Fusion Network (ResCrossNet) for hyperspectral deforestation change detection. The model processes a pair of hyperspectral images captured at different time intervals over the same geographical area, using identical spatial-spectral feature extraction pipelines. The Spatial-Spectral Feature Extraction Module (SSFM) incorporates residual spatial and residual channel attention blocks to effectively capture meaningful spatial and spectral information. These extracted features are then subjected to the Cross-Fuse Module (CFM), which increases complementary features and reduces redundancy between the bi-temporal inputs. The fused features are integrated using differencing and concatenation techniques to detect both obvious and subtle changes in the forest landscape. Finally, semantic segmentation is performed using the Refined Object-Contextual Representation (ROCR) module, which models object-level contextual dependencies to increase the accuracy of pixel-level change. Its excellent categorization accuracy of 99.6% when implemented with Python showed its potential for accurate and scalable monitoring of forest change.