Earthquakes often lead to significant secondary hazards such as landslides, liquefaction, and aftershocks, which in turn cause great damage to buildings and seriously jeopardize socio-ecological welfare. The prevailing models for post-earthquake damage assessment predominantly utilize deep learning methods and InSAR-based Damage Proxy Maps. However, these approaches require data of high quality, with both multi-temporal and spatiotemporal resolution, and are heavily reliant on supervised learning, limiting their applicability on a broader scale. This paper presents a Case-Based Reasoning Framework Augmented with Causal Graph Bayesian Networks for Multi-Hazard Impact Assessment. This method demonstrates strong adaptability to noisy data, making it an innovative tool in the field of earthquake damage estimation. We applied this framework to analyze the catastrophic earthquakes that struck Turkey and Japan in 2023 and 2024, respectively, using them as bases for our case-based reasoning process. For the Turkey case, our model achieved a precision of 99.9%, a recall of 40.2%, and an F1 score of 57.4% in detecting landslides-significantly surpassing the performance of the USGS a priori model. In detecting liquefaction, the model showed a recall of 95.9% and an F1 score of 70.6%, both substantial improvements over the preliminary model. For the 2024 Noto Peninsula earthquake, our method enhanced the Area Under the Curve (AUC) index from 0.73 to 0.77, further validating the efectiveness of our approach. This study ofers a highly precise, scalable, and unsupervised learning method for estimating earthquake disaster damage, providing a valuable asset for optimizing post-disaster resource allocation, reducing economic losses and accurately repairing the environment.
Earthquakes are an unpredictable and devastating disaster that causes significant damage, economic
loss, and risk to life, posing a great challenge to socioeconomic welfare [
To address the above three problems in the estimation of newly occurring large earthquakes, we
introduce the Case-Based Reasoning (CBR) [
Firstly, we collected a large amount of raw data from relevant earthquakes on open-source websites and uniformly converted them into raster data that can be accurately input into a causal Bayesian network for comprehensive validation.
Secondly, we defined conditional probabilities, for each pair of parent and child nodes in the causal
graph, defining the conditional probability relationship between them, and subsequently investigated
the introduction of a variational inference algorithm into the Bayesian network to estimate unobserved
intermediate seismic hazards, building damages, and quantitative causal dependencies between them.
In addition, the system can flexibly integrate available building footprint information provided by
OpenStreetMap or Microsoft to further improve the inference performance of the whole graph [
Finally, by inputting the collated data from the Turkey and Japan earthquakes and using the expectation maximization (EM) algorithm to optimize the model parameters in the causal map, including the causal coeficients and the a posteriori probabilities of the unobserved variables, and by adjusting the hyper-parameters in the network with the characteristics of the actual seismic region, the relevant causal coeficients computed based on the sequence of this earthquake can be obtained and more accurate landslide, liquefaction, and house damage probability.
Theoretical level 1. This study is based on the causal Bayesian network model [
Current methods for assessing earthquake damage predominantly encompass two strategies. The
ifrst leverages multi-temporal radar interferometry to create damage proxy maps (DPMs), which are
instrumental in detecting and annotating changes within remote sensing images of areas afected by
disasters. This approach ofers a rapid estimation of earthquake impacts, exemplified by Yun & Sang-Ho
et al.(2015) [
In pursuit of a more universally applicable approach for large-scale disaster damage assessment, Xu [
Case-based reasoning (CBR) methods, especially those grounded in geographic space, have recently
made significant strides in disaster damage assessment [
The catastrophic earthquakes in Turkey and Japan, occurring in 2023 and 2024 respectively, present a
unique opportunity to refine and apply the space-based CBR method for a comprehensive assessment of
complex earthquake hazards. Drawing inspiration from this research trajectory, we aim to enhance Xu’s
model [
CBR (Case-Based Reasoning) is a problem-solving methodology that leverages past experiences or cases to solve new problems. It operates on the principle that similar problems tend to have similar solutions.
3.1.1. Dataset We mainly selected the 7.8 magnitude earthquake that occurred 23 kilometers east of Nodagi, Gaziantep Province, Turkey in 2023 and the 7.6 magnitude earthquake that occurred in Noto Peninsula, Japan in 2024 for our research and experiments. Building footprint, DPM, Landslides, and Liquefaction, are required to construct the causal Bayesian network. On the open web data platform, we collected a quantity of raw data and constructed it as an original dataset.
DPMs DPMs(Damage Proxy Maps) are tools developed by the NASA Advanced Rapid Imaging and Analysis (ARIA) team for rapidly assessing the extent of damage after disasters. They utilize satellite remote sensing data, particularly radar interferometry data, and optical imagery, to identify and quantify surface damage caused by earthquakes. We collected DPMs of the relevant earthquakes on the oficial ARIA website. Among them, the images of the Turkey earthquake cover part of the seismic region (35.51°N - 38.97°N, 35.58°E - 38.80°E) and are generated at a resolution of 2.8E-4 (latitude/longitude) resolution was generated. Seismic images of the Noto Peninsula in Japan covering part of the area (36.31°N-37.57°N, 136,24°E-137.36°E) were generated at the same resolution.
Landslides and Liquefaction Landslide Maps and Liquefaction Maps are raster data generated by the USGS using known geographic features and ground shaking data, which consisted on Bayesian networks as prior data. We collected landslide and liquefaction image data of the related earthquakes on the oficial website of USGS (earthquake.usgs.gov). In Turkey, Landslide Maps and Liquefaction Maps were generated for some areas (35.06°N -37.07°N:34.47°E -37.15°E) after the earthquake with an accuracy of 16.0 km width of the real area covered by each square pixel element. In Japan, Landslide Maps were generated for some areas (34.75°N - 40.13°N:133.88°E - 140.80°E) with a resolution of 0.002083333, respectively, Landslide maps and liquefaction maps were generated for some areas (34.75°N - 40.13°N, 133.88°E - 140.80°E) in the Noto peninsula region at a resolution of 4.2E-4 (latitude/longitude). Building Footprints Building footprints refer to the outlines or shapes of buildings projected onto the ground. We obtained building footprint images for part of the afected area from the HDX website (data.humdata.org), whose data comes from OSM.
Ground Truth We have collected real earthquake data as test comparisons from the Japan Institute of Land and Geography, GitHub, and other online platforms, including real liquefaction maps of the ground, landslide maps, and building damage maps in the aftermath of the earthquake. 3.1.2. Processing The building footprint, DPM, Landslides, and Liquefaction image data inputted into the Bayesian network must have the same spatial coverage area and precision. Of the raw data, Building Footprints images are in shapefile format and the rest are raster images in tif format. Only when each pixel of the four image datasets is aligned, can the corresponding information on buildings, landslides, liquefaction, etc., be accurately fed into the causal Bayesian network for prediction. Therefore, we primarily conducted a three-step preprocessing on the original dataset: Building Footprints Labeling The original dataset’s building footprint image data was in shapefile format, requiring conversion to raster data for input into the Bayesian network. Firstly, we used ArcGIS to convert the building footprint image to raster data. Then, we assigned values to each pixel in the data, converting it into a binary variable ranging from 0 to 1. Pixels with buildings were marked as 1, while those without were marked as 0.
Image Cropping The spatial ranges of diferent image data in the original dataset varied. The overlapping spatial portion was deemed the efective study area. Thus, we utilized ArcGIS for cropping to ensure that the input building footprint, DPM, Landslides, and Liquefaction image data were of consistent size and spatial overlap.
Precision Alignment (Resampling) The input image data required consistent precision, yet the original dataset’s precision difered among building footprints, landslides, and liquefaction images compared to DPMs, which were coarser. Using the pixel size of DPMs as the reference, we divided the pixel granularity of the other three image datasets to match DPMs’ pixel size. Finally, the precision of the four image datasets was unified to 0.00027777778 degrees latitude/longitude. Thus, we obtained the final dataset, as shown in Table 1. The four image datasets used for model training are unlabeled. Building footprints are binary variables ranging from 0 to 1, while the remaining three datasets are continuous variables. As the degree of damage/liquefaction/landslide increases, the corresponding value approaches 1.
In our problem, Earthquakes can directly trigger secondary disasters such as landslides and liquefaction, which may lead to structural damage and casualties. Concurrently, these secondary disasters, including landslides and liquefaction, can also cause injury to buildings. The complexity is further exacerbated by environmental noise and varying geographical conditions, which complicate the causal relationships between diferent events. In such intricate scenarios, the Bayesian causal network demonstrates its formidable problem-solving capabilities by parsing the intricate causal relationships among a multitude of variables within a vast dataset.
The Causal Bayesian Network [
We used a model based on the one put forward by Xu et al.(2022) [
In our research, We engage with a multitude of stochastic variables, focusing on discerning the causal relationships between the existing coarse-grained landslide and liquefaction probability distributions, which are rich in information and environmental noise, and the joint secondary disasters of landslides, liquefaction, and building damage. This is achieved by leveraging Damage Proxy Maps (DPMs) and local geographical context, thereby synthesizing a joint probability distribution for secondary disasters. Our model incorporates landslide and liquefaction probability maps from the USGS team, DPMs from the ARIA team, and building footprints from OpenStreetMap as inputs. The model then synthesizes and outputs posterior probability maps for landslides, liquefaction, and building damage, as well as causal coeficients between them. The model framework is shown in Figure 2. Building upon this foundation, our model is capable of elucidating complex causal relationships to demystify the mechanisms underlying the occurrence of secondary disasters accompanying earthquakes. It eficiently assesses the primary causes of building damage in specific seismic events and facilitates rapid predictions, all predicated on a rich dataset.
In the specific expression of the formula, We let denote the impact nodes, which include DPMs
(Damage Proxy Maps), LS (Landslides), LF (Liquefaction), and BD (Building Damage). We denote as
the binary variable representing the activity and inactivity of the unobserved ground damage and impact
nodes, where can represent LS, LF, BD here, ∈ {1, 2, 3}and is a binary variable with ∈ {0, 1}.
Outside of the aforementioned definition, a bias node 0 is defined as 0 = 1, indicating that it is always
active regardless of the activity of its parent nodes, to address certain anomalies, such as instances
where building damage occurs despite the inactivity of parent nodes like landslides and liquefaction.
The parent nodes of the child node are denoted as , where DPMs are denoted separately as .
With the random disturbance term following a normal distribution, then |() follows a log-normal
distribution. When examining the causal influence relationship between parent and child nodes, the
causal influence coeficient of parent node on child node is defined as (), where satisfies [
∑︁ ∈ () ∑︁ ∈ () (1) (2) (3)
To achieve the adaptive application of the model in diferent cases, reasonable parameter estimation and optimization adjustments are required. We iterate and optimize the model parameters by applying the model to existing earthquake cases and comparing them with real data. Furthermore, in the practical application of the model, various adjustments and solutions have been implemented to address specific challenges. As previously mentioned, the intricate causal inferences involve a multitude of factors, including secondary disasters, environmental conditions, and geographical elements. Relying solely on traditional Bayesian Causal Network inference to predict the posterior distributions of secondary disasters is insuficient, as these algorithms may struggle to capture and represent the complexity, particularly when nonlinear relationships among variables are present. Furthermore, the model’s Bayesian Causal Network encompasses a variety of variable types, such as binary, log-normal, and logit-normal, which poses additional challenges for traditional Bayesian algorithms. Moreover, the traditional Bayesian network operates under the assumption of conditional independence among variables, an assumption that is violated in our scenario where secondary disasters exhibit complex and interdependent causal relationships. To address these issues, the model employs a variational inference algorithm that estimates the posterior distributions by maximizing the marginal likelihood of observed variables, a technique proven efective for complex probabilistic models with numerous unobserved variables. By maximizing the variational lower bound, the algorithm optimizes the posterior distributions of location-specific multi-hazard and impact variables along with their causal dependencies, providing a theoretical guarantee for the optimality of joint posterior inference. Specifically, for each location , we define a variational distribution () to further decompose the unobserved variables: () = ∏︁ () = ∏︁() (1 − )1− (4) + − ∈ () ∑︁ ∈ () ∈ {1,2,3} 2 ∑︀∈ (,)
22 By maximizing the lower bound, we express the optimal combination of expected posterior estimation and causal relevance estimation. Define a nonlinear function: (.) = ∑︁ {(− 1)(0 + + 22 ) + (− 1)(1 − ) ((− 1)0 + 22 )} {(− 1) ((− 1)(0 + ) + 22 ) + (− 1)(1 − ((− 1)0 + 22 )} (5) + (2 − 20 − 2)()
22
where (), , and p(i) represent the parent nodes, c(i) represents the child nodes, and s(i) represents
the set of spouse nodes sharing the same child nodes. Based on this, we provide the optimal posterior
estimation [
Recognizing that complex causal relationships do not significantly manifest in all regions, the model leverages this insight to enhance computational eficiency, particularly valuable for post-earthquake loss assessment where time is of the essence. In areas devoid of building footprints or where the prior probability of a single disaster is low, local pruning strategies can be employed to designate certain nodes as inactive, efectively setting their posterior probabilities to zero. This approach streamlines the impact estimation process. To strike a balance between model complexity and predictive accuracy, it is imperative to ensure that the model’s accuracy remains above a performance benchmark. This is achieved through gradual pruning and iterative model adjustments.
By applying these methods, the model can be extended to various types of disasters, ofering significant practical value that could be proved in the results. With a rich case library formed from a large number of earthquake cases and the adaptability training of the model to diferent earthquakes, we can efectively obtain probability estimates for various specific seismic losses in each location. The model’s adaptability and eficiency, coupled with its ability to maintain high predictive accuracy, make it a robust tool for disaster risk analysis across diverse seismic environments.
Our model has output three probability maps for landslides, liquefaction, and building damage, visually presenting the likelihood of these damages occurring at diferent locations with varying shades of color depth. We compared these probability maps with the actual damage maps obtained after the earthquake to assess the performance of the model. On the other hand, we demonstrated the refinement of our model’s results by comparing them with the landslide and liquefaction probability maps from the USGS as prior information in the input data. From both evaluations, we achieved favorable outcomes.
It is important to evaluate the accuracy of this multi-hazard fast response model for damage prediction of newly occurring earthquakes, as well as to compare the performance improvement of the prior model. We summarize the performance of our system and prior methods using the model’s Precision, Recall, F1 score, and ROC curves, respectively. Table 2 shows the results of our test of causal Bayesian networks on the 2023 Turkey earthquake dataset.
We evaluate the predictions of our system using ground-truth data collected by a post-event reconnaissance team and compare the prediction performance with existing USGS earthquake ground damage models (i.e., prior models for landslides, and liquefaction). For landslide predictions, our model predicted a Precision of 99.9%, Recall of 40.2%, and F1 score of 57.4%, which is a significant increase compared to the USGS a priori model (Nowicki Jessee, 2018), and we predicted landslides very accurately while predicting more imminent areas of landslides. In the prediction of liquefaction, our model has a recall of 95.9% and an F1 score of 70.6%, both of which are much improved over the a priori model, but the accuracy still needs to be improved. Overall, we accurately predicted most of the areas where liquefaction is imminent.
Also, we summarize the performance of our system and a priori method for building damage prediction using a subject operating characteristic (ROC) curve which visualizes the variation of true positive rate versus false positive rate. Figure 4 gives the performance evaluation results for building damage in the Turkey earthquake, and it can be seen that we have high accuracy in predicting building damage in the Turkey earthquake. Figure 5 gives the performance evaluation results for landslide prediction in the Noto Peninsula earthquake in Japan, and the AUC of our model is 0.77, which improves the performance of our model by 5.36% compared to the USGS model. In summary, our model efectively integrates information from building footprints, a priori models, and DPMs, which significantly improves the performance of ground fault estimation compared to traditional geospatial models and models based solely on DPMs.
4.2.1. 2023 Turkey earthquake We analyzed the causal dependencies between ground shaking, landslide, liquefaction, and building damage using a weight matrix. We found that in the southwestern region, building damage is mainly caused by liquefaction rather than landslides. In earthquakes in Turkey, the causal coeficient from liquefaction to building damage in the southwest is 3.4078, significantly higher than the coeficient from landslides to building damage, which is 0.0301. In the central-northern region, building damage is caused by the combined efects of landslides and liquefaction, with coeficients for liquefaction and landslides at 1.9050 and 1.9021 respectively.
Soil liquefaction refers to the significant reduction in strength and stifness of saturated or partially saturated soil lacking cohesion in response to ground motion during earthquakes, essentially making the soil behave like a liquid. Therefore, liquefaction occurring beneath buildings and other structures can lead to significant damage, including severe tilting of buildings, ground subsidence, and lateral flow of soil during intense seismic events. In the 7.8 magnitude earthquake in Turkey, extensive liquefaction phenomena were observed in Gölbaşı (Adıyaman) by the lakeshore, the İskenderun port area, and along the Aşu River near Antakya. These areas are located in the southwestern part of Turkey, consistent with our model results. In the central-northern regions of the study area, which are mostly mountainous, landslides also become significant factors afecting building damage, consistent with the model results. 4.2.2. 2024 Japan earthquake In the Japanese earthquake of January 2024, we found that building damage was caused by both landslides and liquefaction, with a causal coeficient of 0.7125 from liquefaction to building damage and 0.7105 from landslides to building damage. Since there was no significant diference between the two coeficients, we conclude that building damage in the Japanese earthquake resulted from the combined efects of these two factors.
Around 1:10 AM on January 1, 2024, a magnitude 7 earthquake centered in Suzu City, Prefecture, struck, causing severe devastation across the Noto Peninsula. Consequently, even in the central regions, there was significant slope degradation. This demonstrates consistency between landslide and liquefaction-induced building damage and the model results.
The United States Geological Survey (USGS) has developed a series of global earthquake monitoring and rapid assessment systems, including ShakeMap, Did You Feel It? (DYFI), Earthquake Early Warning (EEW), and Prompt Assessment of Global Earthquakes for Response (PAGER), all of which have reached the state-of-the-art (SOTA) level, but each has certain drawbacks. ShakeMap provides earthquake intensity maps to assist decision-makers and emergency response agencies in quickly understanding the extent of earthquake impact in diferent areas; however, its accuracy and reliability are constrained by the seismic monitoring network, and it only provides earthquake intensity assessments rather than comprehensive earthquake disaster assessments. The DYFI system relies on subjective reports from the public, which may introduce subjectivity and uncertainty, and may lack suficient quantity and quality of reports in some areas, afecting the accurate assessment of earthquake impact. The EEW system provides limited warning time, possibly only a few seconds to tens of seconds, and is susceptible to false alarms and missed alerts, reducing people’s trust in the system. The PAGER system aims to provide rapid assessment after earthquakes, but still faces certain delays and accuracy issues, especially in the early stages following an earthquake event.
Our model, leveraging Damage Proxy Maps (DPM), ofers several advantages over existing earthquake monitoring and assessment systems. By utilizing DPM, we have achieved higher-resolution disaster predictions with enhanced accuracy and speed of response. Here are the key benefits of our model: 1. Improved Resolution: Our model provides higher-resolution predictions compared to existing systems, allowing for more detailed and localized assessments of earthquake impacts. This finer granularity enables decision-makers and emergency response agencies to better allocate resources and prioritize areas for assistance. 2. Enhanced Accuracy: Leveraging advanced probabilistic modeling techniques through DPM, our model delivers more accurate assessments of earthquake intensity and potential damage. By incorporating a wider range of data sources and advanced algorithms, we minimize the limitations associated with traditional seismic monitoring networks, resulting in more reliable predictions. 3. Rapid Response: Our model enables swift response to earthquake events by providing timely and actionable information to relevant stakeholders. The speed of our system’s predictions, facilitated by DPM, allows for quicker mobilization of resources and implementation of emergency measures, potentially reducing the impact of earthquakes on afected communities.
Our method, which takes into account the complex interplay of multiple factors post-earthquake, provides a joint estimation of multiple seismic disaster losses. Quantifying the various triggering factors that cause building damage ofers significant assistance for actual post-disaster response scenarios. Notably, the application of variational inference and local pruning ensures computational eficiency and scalability, efectively utilizing information from DPMs without waiting for ground truth data, thus accelerating the process of obtaining results. This enables the model to rapidly analyze and estimate losses immediately after an earthquake, making it suitable for various real-world scenarios. In earthquake disaster events, ground motion has always been a critical factor, either directly or indirectly, leading to building damage. However, there are also complex causal relationships between the various post-earthquake events triggered by ground motion. In the February 2023 Turkey earthquakes, the results indicate that the most significant factor afecting building damage was landslides. This observation is consistent with the extensive landslide phenomena noted in the severely impacted regions within the vicinity of our study area. Similarly, In the January 2024 earthquake in Japan, the model results indicate that the building damage caused by both landslides and liquefaction is consistent with the real data. The unique circumstances present in diferent earthquakes can greatly afect the pattern of loss formation following the seismic event. Rapidly analyzing and estimating these conditions is the central aim of this research. Based on a multi-layer Bayesian causal network, this loss estimation method can be extended to loss estimation for a broader range of catastrophic events. By transforming heterogeneous factors such as disasters and their induced secondary hazards and impacts into nodes within a Bayesian physical causality network, this approach integrates complex information into a holistic model connected by causal relationships.
This work was jointly supported by National Natural Science Foundation of China (NSFC) under grants 62206301; JSPS KAKENHI (Grants-in-Aid for Scientific Research, 21H05001); Public Health & Disease Control and Prevention, Fund for Building World-Class Universities (Disciplines) of Renmin University of China. Project No. 2024PDPC; the Major Project of the MOE (China) National Key Research Bases for Humanities and Social Sciences (22JJD910003); and this research was supported by Public Computing Cloud, Renmin University of China. We sincerely thank Dr. Zuo Zhenpeng of Boston University for providing data processing support.