ADS Capstone Chronicles Revised

14

overfitting, leading to an output layer with softmax activation for multi-class classification. This architecture effectively combines transfer learning with task-specific enhancements for robust feature learning.

Furthermore, “the IoU is calculated by taking the intersection area of two bounding boxes and dividing it by the union area of these boxes” (Vedoveli, 2023). MSE measures the difference between predicted and ground truth values, indicating how well the model performs overall. Confusion matrix provides insights into true positives, false positives, false negatives, and true negatives, helping evaluate classification performance, while robustness accuracy and robustness loss reflect the model’s generalization and stability across varying conditions. Building localization in pre-disaster images necessitates high-performance semantic segmentation models that accurately identify building regions while balancing computational efficiency and precision. Two state-of-the-art models — FCN and U-Net — were implemented and evaluated based on their architecture and segmentation performance metrics. As shown in Table 2, the FCN demonstrated strong overall performance, achieving a precision of 0.94 and an F1-Score of 0.91, reflecting its ability to identify building regions accurately while maintaining a balance between precision and recall. Its IoU of 0.84 and pixel accuracy of 0.98 indicate that the model is effective in localizing buildings but occasionally misses finer details, as shown by its higher false negatives count. The U-Net model excelled in precision delineation of object boundaries, achieving a higher Recall of 0.96 and an F1-Score of 0.93, which highlights its robustness in capturing subtle details and ensuring fewer missed building regions, as reflected by its lower false negatives count. With an IoU of 0.88 and pixel accuracy of 0.99, U-Net provides superior segmentation performance, benefiting from its skip connections that preserve spatial details across encoder-decoder layers. This architecture also yielded a lower MSE (0.01), 5.2 Building Localization Result

5 Results and Findings

5.1 Image Segmentation Metrics

In the task of building localization in pre-disaster images and damage classification post-disaster images, high-performance models are critical for accurate segmentation and identification of building regions. For classification tasks, metrics such as precision, recall, F1-score, intersection over union (IoU), pixel accuracy, and mean squared error (MSE) were used to evaluate the models' ability to detect and segment building localization and damage. Precision and recall measure the model's ability to correctly identify damage and capture all relevant damage incidents. F1-score combines these two to provide a balanced view of the model’s performance. IoU evaluates the overlap between predicted and actual building regions, while pixel accuracy calculates the ratio of correct pixels to total pixels, ensuring that the model accurately identifies each pixel in the image. Pixel accuracy is a simple metric, but it can be misleading in cases where the dataset is imbalanced (Geeksforgeeks, 2024).

Figure 5 Intersection over Union

280