Scholarlink

<p>Land use/Land cover (LULC) maps provide valuable information about the Earth’s surface and are crucial for various disciplines, including urban planning, environmental management, agriculture, and climate studies. However, the classes that represent the earth’s surface are quite imbalanced as they are underrepresented. This imbalance often leads to biased learning algorithms that favor the majority class. To address this problem, data augmentation is commonly employed to rebalance the data set by introducing minority class samples. This research paper proposes a novel multi-class oversampling approach using deep generative adversarial networks (MO-DGAN) to tackle the class imbalance problem. In contrast to traditional GAN models that focus on a single class, the proposed oversampling technique integrates multiple minority classes using a class-conditioned Variational Autoencoder, to capture spectral and spatial features of the image and learn the distribution in the latent space, a modified generator with an embedded layer and decoder to generate minority class samples. The classifier was chosen following a comprehensive comparative analysis of five state-of-the-art CNN architectures. Among them, ResNet101 demonstrated the best performance and was subsequently employed as a discriminator to classify images into their respective classes and distinguish between real and synthetic samples. These components work together to generate high-quality, diverse minority class samples that significantly enhance classification performance. Extensive experiments are conducted using publicly available remote sensing datasets, i.e., Eurostat. In comparison to a state-of-the-art GAN model, the experimental results demonstrate that MO-DGAN may produce different instances of minority classes, even when the quantity of such instances is constrained. Moreover, our model performs better in terms of classification accuracy and the caliber of the produced samples than other widely used augmentation and oversampling techniques.</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

Multi-class oversampling approach using deep generative adversarial network (MO-DGAN) for remote sensing imagery

  • 期刊论文

  • 发表日期：

    2025年12月9日

查看全文
  • Shweta Sharma,
  • Anjana Gosain

摘要

Land use/Land cover (LULC) maps provide valuable information about the Earth’s surface and are crucial for various disciplines, including urban planning, environmental management, agriculture, and climate studies. However, the classes that represent the earth’s surface are quite imbalanced as they are underrepresented. This imbalance often leads to biased learning algorithms that favor the majority class. To address this problem, data augmentation is commonly employed to rebalance the data set by introducing minority class samples. This research paper proposes a novel multi-class oversampling approach using deep generative adversarial networks (MO-DGAN) to tackle the class imbalance problem. In contrast to traditional GAN models that focus on a single class, the proposed oversampling technique integrates multiple minority classes using a class-conditioned Variational Autoencoder, to capture spectral and spatial features of the image and learn the distribution in the latent space, a modified generator with an embedded layer and decoder to generate minority class samples. The classifier was chosen following a comprehensive comparative analysis of five state-of-the-art CNN architectures. Among them, ResNet101 demonstrated the best performance and was subsequently employed as a discriminator to classify images into their respective classes and distinguish between real and synthetic samples. These components work together to generate high-quality, diverse minority class samples that significantly enhance classification performance. Extensive experiments are conducted using publicly available remote sensing datasets, i.e., Eurostat. In comparison to a state-of-the-art GAN model, the experimental results demonstrate that MO-DGAN may produce different instances of minority classes, even when the quantity of such instances is constrained. Moreover, our model performs better in terms of classification accuracy and the caliber of the produced samples than other widely used augmentation and oversampling techniques.