ADS Capstone Chronicles Revised

16

indicating superior segmentation capabilities. Additionally, its pixel accuracy of 0.7034 which was the highest among the models, demonstrating its ability to correctly identify the relevant areas in the images. The model also exhibited the best robustness accuracy (0.5513), showcasing its ability to generalize well in varying conditions. Table 3 Damage Classification Performance Metrics

ability to not only classify damage accurately but also segment the damaged areas effectively. MobileNet’s efficiency in handling resource constraints, coupled with its ability to generalize well across different scenarios, makes it the most reliable choice for this task. While SimpleCNN and ResNet50 also delivered solid results, MobileNet’s balanced combination of accuracy, efficiency, and robustness positions it as the optimal model for real-world deployment in building damage detection applications. Due to its strong performance metrics, MobileNet's lightweight architecture makes it particularly advantageous for this real-time application in disaster response scenarios. Its ability to operate effectively under resource constrained environments solidifies that it can deliver quality results without requiring extensive computational power. Having this flexibility along with its capability is especially vital in emergency situations where timely and reliable damage assessments are paramount. This model achieves a complimentary balance between computational efficiency and high accuracy. MobileNet, not only outperforms more complex models like ResNet50 but also demonstrates its practical value in scaling solutions for large-scale disaster management efforts. The adaptability is important especially across diverse disaster scenarios. Ultimately, as a result, MobileNet emerges as a potentially robust choice for streamlining post-disaster recovery initiatives and driving informed decision-making.

Model

Simple CNN

MobileN et

ResNet50

Precision

0.5522 0.7041

0.6738

Recall

0.5814 0.7034

0.6674

F1-Score

0.5521 0.7029

0.6694

IoU

0.4001 0.5461

0.5066

Pixel Accuracy

0.5814 0.7034

0.6674

MSE

0.6954 0.5325

0.6497

True Positives

4207

5090

4829

False Positives

3029

2146

2407

False Negatives

3029

2146

2407

True Negatives

18679 19562

19301

Robustness Loss

0.3011 0.1944

0.2266

Robustness Accuracy 0.2282 0.5513

0.4502

The results from ResNet50, while promising, showed it to be less effective in this task compared to MobileNet, specifically in terms of precision, recall, and robustness. MobileNet's overall performance indicates its strong balance between accuracy, efficiency, and generalization, making it the most suitable model for building damage classification. Its superior performance in key metrics such as precision, recall, F1-score, IoU, pixel accuracy, and robustness highlights its

282