Scholarlink

<p>The proposed post-disaster damage classification using satellite images aims to analyse building damage after natural disasters and classify it into four categories, such as No Damage, Minor Damage, Major Damage, and Destroyed. The proposed system applies image enhancement techniques to improve contrast and edge visibility. The satellite images are then passed to a U-Net–based segmentation model to localize buildings. The U-Net model utilises an Efficient- NetV2 as backbone. The classification stage employs a Siamese neural network initialized with weights transferred from the U-Net model. The pre-disaster and post-disaster image pairs are fed to a classification model to classify the buildings in the images based on the level of damage. The performance is evaluated based on F1-score and Dice coefficient, owing to class imbalances. The proposed approach, EfficientU-NetV2 achieves a Dice score of 0.78 and an F1-Score of 0.60 in the localisation task, and a Dice score of 0.60 and an F1-score of 0.525 in the classification task. The weighted F1-score is used to assess the performance of the localisation and classification model as a whole. The proposed system achieves a weighted F1-score of 0.55, which is significantly higher than the baseline model’s weighted F1-score.</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

Deep learning based post-disaster damage classification using satellite images

  • 期刊论文

  • 发表日期：

    2026年4月8日

查看全文
  • Valliyammai C,
  • Vaishnavi A. K,
  • Surya Siddharthan S,
  • Aravind Priyamvadan V,
  • Dhilip Kumar G

摘要

The proposed post-disaster damage classification using satellite images aims to analyse building damage after natural disasters and classify it into four categories, such as No Damage, Minor Damage, Major Damage, and Destroyed. The proposed system applies image enhancement techniques to improve contrast and edge visibility. The satellite images are then passed to a U-Net–based segmentation model to localize buildings. The U-Net model utilises an Efficient- NetV2 as backbone. The classification stage employs a Siamese neural network initialized with weights transferred from the U-Net model. The pre-disaster and post-disaster image pairs are fed to a classification model to classify the buildings in the images based on the level of damage. The performance is evaluated based on F1-score and Dice coefficient, owing to class imbalances. The proposed approach, EfficientU-NetV2 achieves a Dice score of 0.78 and an F1-Score of 0.60 in the localisation task, and a Dice score of 0.60 and an F1-score of 0.525 in the classification task. The weighted F1-score is used to assess the performance of the localisation and classification model as a whole. The proposed system achieves a weighted F1-score of 0.55, which is significantly higher than the baseline model’s weighted F1-score.