Scholarlink

<p>Evaluating damage from buildings in the timeliest and most accurate manner is critical for response following a disaster. Here we assess the performance effectiveness of a pre-trained Vision Transformer (ViT-B32) relative to a standard Convolutional Neural Network (EfficientNet-B0) binary classification scheme for damaged vs. undamaged buildings based on 224 × 224 RGB post-hurricane satellite imagery (Hurricane Harvey, August 2017, Texas, USA; 10000/2000/2000 balanced test and 9000 unbalanced test images), obtained from a public benchmark dataset. In contrast, ViT-B32 (input patch size: 32 × 32, embedding dimension: 768, number of transformer layers: 12 and pretrained on ImageNet-21k with 21,841 classes) has the added ability to encode global spatial relationship using self-attention mechanisms a property we considered critical for differentiating between subtle damage features from text that is predominantly visually cluttered as it is in disaster zones where this task directly apply. In the experiments, model ViT shows better classification accuracy (97.85% against 96.90%) and is more robust on unbalanced data in comparison with respective close competitors while using F1-score and Area under ROC curve as metrics. So, we interpret this model by creating attention heatmaps showing what parts of the image drove classification. These visualization’s not only aids in actionable information, specifically accurately locating where the structural damage has occured but also plays a pivotal role in recovery oriented disaster management workflows; the statistically significant performance gain (χ² = 772.15, <i>p</i> &lt; 0.001) on both balanced and unbalanced test sets signals this could provide to be a valuable asset to field of disaster classification methods. Overfitting was handled by (i) freezing the backbone in Stage 1 (therefore preventing any gradient update to ~ 86.6&#xa0;M parameter model); (ii) applying Dropout(0.2) on the classification head; and (iii) early stopping based on validation AUC.</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

Vision transformer based damage assessment from post-disaster satellite imagery: an applied study on hurricane harvey

  • 期刊论文

  • 发表日期：

    2026年5月21日

查看全文
  • Jakaria Habib,
  • Imran Hossain,
  • Basim Al Zabir Shammo,
  • Taz Warul Mulk,
  • Estiyak Rubaiat

摘要

Evaluating damage from buildings in the timeliest and most accurate manner is critical for response following a disaster. Here we assess the performance effectiveness of a pre-trained Vision Transformer (ViT-B32) relative to a standard Convolutional Neural Network (EfficientNet-B0) binary classification scheme for damaged vs. undamaged buildings based on 224 × 224 RGB post-hurricane satellite imagery (Hurricane Harvey, August 2017, Texas, USA; 10000/2000/2000 balanced test and 9000 unbalanced test images), obtained from a public benchmark dataset. In contrast, ViT-B32 (input patch size: 32 × 32, embedding dimension: 768, number of transformer layers: 12 and pretrained on ImageNet-21k with 21,841 classes) has the added ability to encode global spatial relationship using self-attention mechanisms a property we considered critical for differentiating between subtle damage features from text that is predominantly visually cluttered as it is in disaster zones where this task directly apply. In the experiments, model ViT shows better classification accuracy (97.85% against 96.90%) and is more robust on unbalanced data in comparison with respective close competitors while using F1-score and Area under ROC curve as metrics. So, we interpret this model by creating attention heatmaps showing what parts of the image drove classification. These visualization’s not only aids in actionable information, specifically accurately locating where the structural damage has occured but also plays a pivotal role in recovery oriented disaster management workflows; the statistically significant performance gain (χ² = 772.15, p < 0.001) on both balanced and unbalanced test sets signals this could provide to be a valuable asset to field of disaster classification methods. Overfitting was handled by (i) freezing the backbone in Stage 1 (therefore preventing any gradient update to ~ 86.6 M parameter model); (ii) applying Dropout(0.2) on the classification head; and (iii) early stopping based on validation AUC.