Scholarlink

<p>Remote sensing scene classification (RSSC) faces two significant challenges: class imbalance and interclass similarities between the scenes. Despite the methodological advancement of deep learning algorithms in RSSC, there are still notable research gaps that need to be addressed. Firstly, existing deep learning techniques primarily focused on feature space modifications, which often struggle if classes exhibit poor linear separability. This motivates the need to design algorithmic based classifiers to address class imbalance at the classification stage. Secondly, transferring feature statistics from majority to minority through reweighting or resampling can introduce bias in classifier learning. Thirdly, the limitation of CNN-based architectures is that they do not capture the feature interdependencies between the layers that can further enhance the accuracy of similar scenes. The cross-entropy (CE) loss function gives suboptimal performance in imbalanced datasets and for classes exhibiting similarities. This paper proposes a Scene-Net framework to address the highlighted issues to minimize misclassifications in RSSC tasks. Specially Cost-aware Focal Hinge loss function that adaptively adjusts the loss contribution based on the classification difficulty of scenes with margin based learning. An SVM classifier with CAFHL loss (CAFHL-SVM) is proposed that iteratively adjusts the weights to shift the attention of the classifier toward scenes with low classification probability. The proposed Scene-Net framework incorporates a CNN-based feature extraction pipeline with multi-level feature fusion followed by the proposed CAFHL-SVM classifier. Extensive experiments on the Eurosat and UCM datasets in different imbalance settings demonstrate the consistent performance of the proposed CAFHL-SVM classifier over baseline weighted losses on key metrics G-Mean, minority class recall, and Macro F1-score. Further, probability studies and confusion matrix analysis indicate its effectiveness for classifying visually similar classes.</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

Scene-net framework for remote sensing scene classification using cost-aware focal hinge loss-SVM classifier

  • 期刊论文

  • 发表日期：

    2025年6月19日

查看全文
  • Neha Kumari,
  • Sonajharia Minz

摘要

Remote sensing scene classification (RSSC) faces two significant challenges: class imbalance and interclass similarities between the scenes. Despite the methodological advancement of deep learning algorithms in RSSC, there are still notable research gaps that need to be addressed. Firstly, existing deep learning techniques primarily focused on feature space modifications, which often struggle if classes exhibit poor linear separability. This motivates the need to design algorithmic based classifiers to address class imbalance at the classification stage. Secondly, transferring feature statistics from majority to minority through reweighting or resampling can introduce bias in classifier learning. Thirdly, the limitation of CNN-based architectures is that they do not capture the feature interdependencies between the layers that can further enhance the accuracy of similar scenes. The cross-entropy (CE) loss function gives suboptimal performance in imbalanced datasets and for classes exhibiting similarities. This paper proposes a Scene-Net framework to address the highlighted issues to minimize misclassifications in RSSC tasks. Specially Cost-aware Focal Hinge loss function that adaptively adjusts the loss contribution based on the classification difficulty of scenes with margin based learning. An SVM classifier with CAFHL loss (CAFHL-SVM) is proposed that iteratively adjusts the weights to shift the attention of the classifier toward scenes with low classification probability. The proposed Scene-Net framework incorporates a CNN-based feature extraction pipeline with multi-level feature fusion followed by the proposed CAFHL-SVM classifier. Extensive experiments on the Eurosat and UCM datasets in different imbalance settings demonstrate the consistent performance of the proposed CAFHL-SVM classifier over baseline weighted losses on key metrics G-Mean, minority class recall, and Macro F1-score. Further, probability studies and confusion matrix analysis indicate its effectiveness for classifying visually similar classes.