Automatic post-tsunami loss modeling using deep learning CNN case study: Miyagi and Fukushima Japan tsunami

Assessing the destruction caused by a tsunami is a challenging task that must be completed quickly with limited resources and information. To address this issue, we propose a method for accurate damage mapping using binary classification of high-resolution satellite imagery, where we enhance the performance of three pre-trained deep neural network models (Vgg19, Inception, and Xception). The pre-trained models are used, which have been previously trained on large datasets, and transferred to our tsunami problem. We also develop a custom network architecture specifically designed for tsunami damage detection using high-resolution remote sensing data, improving the accuracy of automated binary classification. We investigate the impact of various parameters and learning rates to detect small objects, demonstrating the suitability of our approach for tsunami damage assessment. Our network outperforms traditional and current deep learning-based approaches, as it shows low bias and high variance datasets that result in a skillful model. Specifically, we observe that Inception-v3 performs best on the dataset, exhibiting good behavior with low errors and achieving the best overall score with 24.11 min, while other models score between 30.50 min for Vgg19 and 45.33 min for Xception. Our study focuses on two important binary classification categories, tsunami-stricken and non-stricken areas, for which we train the proposed framework on a dataset comprising 30,000 small tiles of high-resolution satellite images obtained from Mexer satellite images. The model is validated on 8000 images using the Jupyter notebook of the Anaconda deep learning framework.