Sesa Wiguna

Owing to recent extreme weather events, flood risk has been rising annually, increasing the demand for fast and accurate flood mapping. Synthetic aperture radar imagery has received considerable attention for flood-mapping applications owing to its all-weather, day-and-night imaging capabilities. Although previous studies have achieved accurate mapping in non-urban areas, challenges remain for urban regions. This study focuses on flood events in Japan by employing a deep learning model and PALSAR-2 imagery to classify non-flooded areas, floods in open areas, and floods in urban areas. To understand the complex spectral characteristics specific to urban areas, this study investigates the integration of geographical features, such as slope and building footprints, into the segmentation process. The experimental results suggest that the inclusion of these supplementary data improves the prediction performance of the trained models.

Deep learning models are generally less able to maintain their performance in out-of-domain testing. Model transferability is crucial, especially when a model needs to be applied to a new dataset, such as in disaster emergency response, where the training samples are scarce. To solve the aforementioned issues, we propose a semi-supervised framework to improve model generalization by utilizing unlabelled samples from the target domain. The framework consists of two main steps: model initialization, which incorporates past events, and iterative fine-tuning. The latter step relies heavily on the pseudo-labels inferred with high confidence from the former step. We tested our framework on the 2024 Noto Peninsula Earthquake. Our framework shows an improvement in the model generalization indicated by higher scores in the tuned model compared to the initial model. The effect is even greater when the local context from the past event is included in the initial learning step. In this case, the score has increased by about 21% from 0.62 to 0.75. The proposed framework offers a promising solution for rapid disaster damage mapping.

Integrated with remote sensing technology, deep learning has been increasingly used for rapid damage assessment. Despite reportedly having high accuracy, the approach requires numerous samples to maintain its performance. However, in the emergency response phase, training samples are often unavailable. Since no ground truth data is available, deep learning models cannot be trained for this specific situation and thus have to be applied to unseen data. Previous research has implemented transfer learning techniques to solve data unavailability. However, many studies do not accurately reflect the rapid damage mapping in real-world scenarios. The present study illustrates the use of earth observation and deep learning technologies in predicting damage in realistic emergency response settings. To this aim, we conducted extensive experiments using historical data to find the best model by examining multiple Neural Network models and loss functions. Then, we evaluated the performance of the best model for predicting building damage due to two different disasters, the 2011 Tohoku Tsunami and the 2023 Turkiye-Syria Earthquake, which were independent of the training samples. We found that a Transformer-based model with a combined Cross Entropy Loss and Focal Loss generates the highest scoring values. The testing on both unseen sites illustrates that the model can perform well in no-damage and destroyed classes. However, the scores dropped in the middle class. We also compared our transformer-based approach with other state-of-the-art models, specifically the xView-2 winning solution. The results show that the transformer-based models have stable generalization toward multi-class classification and multi-resolution imagery.