Ruben Vescovo

Tsunamis, while infrequent, can be some of the most damaging natural hazards to coastal communities. The unpredictable nature of tsunamis necessitates the development of robust methods that quantify probable damage and inform the development of mitigation strategies. While this critical need is recognized by the scientific community, probabilistic frameworks to assess risk given the range of tsunami hazards are still in their infancy. Tsunami Fragility Functions represent the state-of-the-art for building fragility estimation, they link hazard to risk, and are central to these frameworks. These models provide a notion of the expected building damage given a measure of the tsunami. However, they have seldom been studied directly as building damage estimators, hence their capacity to inform downstream tasks remains largely untested. Furthermore, the current state-of-the-art lacks a method to handle input and output uncertainty, which is crucial for holistic frameworks such as recently proposed probabilistic Hazard-to-Risk frameworks. In the present work we propose a novel bayesian ordinal regression framework, inspired by fragility functions, for disaggregated building damage estimation to fulfill this need. We fit our method on 2011 Tohoku Earthquake and Tsunami data and provide three case studies to highlight the model’s generalization capabilities and limitations.

We present a building damage dataset following the 2024 Noto Peninsula Earthquake. The database was compiled from freely available, multi-source, remote sensing data, verified through opt-in crowd-sourced information. The dataset consists of geo-referenced vector polygons representing the pre-event building footprints of 140,208 structures. Each building was classified through visual inspection using pre-disaster and post disaster vertical, oblique, survey, and verifiable news reporting imagery. Entries were validated using voluntary-submission data sourced through a web-API hosting a live version of the database. We calculate classification metrics for a subset of the database where ground survey photographs were provided by independent surveyors. An average F₁-score of 0.94 suggests that the proposed assessment is consistent and high quality. We aim to inform future disaster research such as disaster dynamics models; statistical and machine learning damage models; logistics and evacuation studies. The present work describes the data collection process, damage assessment methodology, and rationale; including limitations encountered, the crowd sourcing validation process, and the dataset structure.

The tsunami was generated by the Mw7.6 Noto Peninsula Earthquake and left widespread impact. After the event occurred, we modeled the tsunami propagation and coastal inundation with various tsunami source models and discussed its propagation and inundation features.Preliminary tsunami modeling results imply that the impacts were severe around Noto Peninsula (Shika to Nanao). Specific bathymetric features of the continental shelf of Noto Peninsula were responsible for high tsunamis in Suzu City. The directivity of tsunami energy was also toward the Japan Sea coasts, especially Joetsu City, Nigata Prefecture. Early tsunami arrival at Toyama City with the leading negative wave could not be explained by fault rupture. The post-tsunami field survey teams at Suzu City preliminarily found tsunami run-ups of 3 m or higher with flow depths of 2.5m or higher. Inside the tsunami inundation zone around Noto Peninsula, we found at least 648 houses were destroyed by both the strong ground motion and tsunami.

The tsunami generated by the Mw7.6 earthquake of Noto Peninsula, Japan left widespread impact. We analyzed multi-modal information and data to elucidate its impact.We modeled the tsunami propagation and inundation with multiple tsunami source models based on GNSS-based crustal movement and tsunami waveform data to understand its propagation and inundation characteristics. The model results are verified by using post-tsunami field survey data. Preliminary tsunami modeling results implied that severe tsunami impacts were around Noto Peninsula (Shika to Nanao). Through the visualization of tsunami propagation model, we found that the remarkable tsunami refraction around the continental shelf of Noto Peninsula were responsible for high tsunamis in Suzu City. This distinctive sea bottom topography also affected the directivity of tsunami energy toward the Japan sea coasts, especially Joetsu city, Nigata Prefecture. Tsunami in Toyama bay had long duration of oscillation caused by multiple-reflection. The leading (negative) tsunami wave could not be explained by fault rupture and this implied the possibility of submarine landslides.The post-tsunami field survey teams at Suzu City preliminarily found tsunami run-ups of 3 m or higher with flow depths of 2.5m or higher. Inside the tsunami inundation zone around Noto Peninsula, we found at least 648 houses out of 3398 were destroyed by both the strong ground motion and tsunami.The cellphone-based population data (Mobile Spatial Statistics) were used to analyze the exposed population in the aftermath of the event. The hourly population estimates with 500m spatial resolution in the coastal communities implied how people reacted and were affected. Approximately 2500 population increase were found in the areas above 10 m after the major tsunami warning was issued.

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.