Scholarlink

Earth observation (EO) data from the Copernicus program has revolutionized the field of remote sensing, providing unprecedented amounts of high-quality satellite imagery. However, extracting meaningful information from this vast data remains a challenge. This study investigates the application of deep learning methods for land use land cover (LULC) classification using a dataset of 213,761 pre-processed Sentinel-2 satellite images representing seven distinct LULC classes in the Indian region. The effectiveness of various semantic segmentation models, including VGG16, Unet + ResNet50, FPN + ResNet50, LinkNet + ResNet50, and Unet + VGG19, was evaluated using both pre-trained ImageNet weights and non-pre-trained weights. The results highlight the challenges in achieving high performance and generalization in satellite image segmentation. While the VGG16 model showed the highest training accuracy, it exhibited significant overfitting and failed to generalize effectively during testing. In contrast, U-Net and ResNet-based models demonstrated more consistent performances across both the training and testing phases. The study emphasizes the need for refining model architectures and exploring different network designs or training regimes tailored specifically to satellite image analysis. Further research should focus on developing architectures and training strategies that cater to the unique characteristics of satellite imagery, such as incorporating domain-specific priors, designing effective data augmentation techniques, and exploring unsupervised and self-supervised learning methods using abundant unlabelled satellite data. Additionally, the development of large-scale, diverse, and well-annotated datasets specific to satellite imagery could significantly enhance the effectiveness of pretraining models and facilitate transfer learning in this field. This research adds to the expanding field of deep learning’s utilization in remote sensing, offering valuable insights into the efficacy of hybrid models and transfer learning for precise and efficient LULC mapping. While deep learning methods show promise for LULC classification using satellite imagery, Additional research and development efforts are necessary to tailor architectures and training strategies specifically for this domain, enabling the complete utilization of deep learning in Earth Observation applications.

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

Comparison of Land Use Land Cover Classification Using Sentinel-2 Images with Different Deep Learning Approaches

  • 章节

  • 发表日期：

    2025年

查看全文
  • Sanskar Jamadar,
  • Shitij Agrawal,
  • Suraj Sawant,
  • Aditya Metha,
  • Amit Joshi

摘要

Earth observation (EO) data from the Copernicus program has revolutionized the field of remote sensing, providing unprecedented amounts of high-quality satellite imagery. However, extracting meaningful information from this vast data remains a challenge. This study investigates the application of deep learning methods for land use land cover (LULC) classification using a dataset of 213,761 pre-processed Sentinel-2 satellite images representing seven distinct LULC classes in the Indian region. The effectiveness of various semantic segmentation models, including VGG16, Unet + ResNet50, FPN + ResNet50, LinkNet + ResNet50, and Unet + VGG19, was evaluated using both pre-trained ImageNet weights and non-pre-trained weights. The results highlight the challenges in achieving high performance and generalization in satellite image segmentation. While the VGG16 model showed the highest training accuracy, it exhibited significant overfitting and failed to generalize effectively during testing. In contrast, U-Net and ResNet-based models demonstrated more consistent performances across both the training and testing phases. The study emphasizes the need for refining model architectures and exploring different network designs or training regimes tailored specifically to satellite image analysis. Further research should focus on developing architectures and training strategies that cater to the unique characteristics of satellite imagery, such as incorporating domain-specific priors, designing effective data augmentation techniques, and exploring unsupervised and self-supervised learning methods using abundant unlabelled satellite data. Additionally, the development of large-scale, diverse, and well-annotated datasets specific to satellite imagery could significantly enhance the effectiveness of pretraining models and facilitate transfer learning in this field. This research adds to the expanding field of deep learning’s utilization in remote sensing, offering valuable insights into the efficacy of hybrid models and transfer learning for precise and efficient LULC mapping. While deep learning methods show promise for LULC classification using satellite imagery, Additional research and development efforts are necessary to tailor architectures and training strategies specifically for this domain, enabling the complete utilization of deep learning in Earth Observation applications.