Scholarlink

<p>The proposed paper aims to address the difficulties of interpreting satellite images by integrating the Segment Anything Model with the U-Net variants, and obtaining better results in terms of quality cloud and land-cover segmenta- tion. This study examines satellite images and best practices to analyze those images. The most difficult task was to acquire the dataset but LSCIDMR has provided us with Satellite images publicly. We performed experiments by com- bining two machine learning models, Segment Anything Model(SAM) and the Attention Mechanisms models. The SAM is used for the Image encoding, and multiple Attention Mechanisms have been separately used to analyze the Images of the Dataset and predicted the pixel accuracy and Other evaluation parame- ters. In this research, the Hybrid SAM model with various attention mechanisms is applied to investigate satellite images. Results show that pixel accuracy has improved. Our proposed hybrid model SAM + SeUNet(Sequeze and Excitation U-Shaped Network) and SAM + UNet(U-shaped Network) have achieved the accuracy of 94.79% &amp; 95.86% respectively outperforming many traditional mod- els. Our research high results show that SAM is emerging technique for the Image segmentation and analysis. This research can be applied in the practical field for the purpose of Weather prediction and to protect ourselves from natural disasters.</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

Data augmentation and normalization strategies for improving satellite image segmentation

  • 期刊论文

  • 发表日期：

    2026年3月7日

查看全文
  • Ayesha Saadia,
  • Muhammad Ishfaq,
  • Muhammad Umair,
  • Iqbal Murtza,
  • Munahil Yaqoob

摘要

The proposed paper aims to address the difficulties of interpreting satellite images by integrating the Segment Anything Model with the U-Net variants, and obtaining better results in terms of quality cloud and land-cover segmenta- tion. This study examines satellite images and best practices to analyze those images. The most difficult task was to acquire the dataset but LSCIDMR has provided us with Satellite images publicly. We performed experiments by com- bining two machine learning models, Segment Anything Model(SAM) and the Attention Mechanisms models. The SAM is used for the Image encoding, and multiple Attention Mechanisms have been separately used to analyze the Images of the Dataset and predicted the pixel accuracy and Other evaluation parame- ters. In this research, the Hybrid SAM model with various attention mechanisms is applied to investigate satellite images. Results show that pixel accuracy has improved. Our proposed hybrid model SAM + SeUNet(Sequeze and Excitation U-Shaped Network) and SAM + UNet(U-shaped Network) have achieved the accuracy of 94.79% & 95.86% respectively outperforming many traditional mod- els. Our research high results show that SAM is emerging technique for the Image segmentation and analysis. This research can be applied in the practical field for the purpose of Weather prediction and to protect ourselves from natural disasters.