Multi-dimensional flood disaster analysis method

Abstract

The invention provides a multi-dimensional flood disaster analysis method. The method comprises disaster space-time analysis, disaster semantic analysis, response adequacy analysis and population exposure degree analysis. According to the disaster situation space-time analysis, disaster situation related information extracted from a text is associated with a time attribute and a space coordinate attribute; analyzing the disaster information from the angles of time sequence and space distribution; the time sequence analysis adopts a quantitative statistics mode, and key nodes of the flood event are analyzed from the precipitation trend, the flood discharge trend, the microblog number trend, the ponding point number trend, the help seeking information number trend and the taken measure number trend; according to disaster semantic analysis, subject analysis is carried out on the text of each help-seeking point, and semantic information of disaster manifestation in space is mined; according to the response measure sufficiency analysis, the disaster response sufficiency is obtained by analyzing the correlation between the number of response measures and the number of help seeking information.

Description

Technical Field

[0001] The present invention relates to the technical field of flood disaster monitoring and disaster analysis, and in particular to a multi-dimensional flood disaster analysis method. Background Art

[0002] Flood disasters have a wide range of impacts on the economy and society, involving multiple dimensions such as infrastructure, social economy, and public health. Therefore, multi-dimensional analysis of flood disasters is of great significance, which can not only fully understand the spatiotemporal evolution characteristics of disasters, but also deeply explore the semantic information and social impact of disasters.

[0003] Spatiotemporal analysis is the basis of flood disaster research. By combining VGI data with traditional observation data to analyze the spatiotemporal distribution and evolution trend of disasters, we can more comprehensively analyze the dynamic evolution of disasters. In recent years, the combination of social media data and remote sensing images has become an emerging direction of spatiotemporal analysis.

[0004] Other studies have shown that social media data has shown good results in flood prediction in basins without monitoring stations. This study demonstrates the spatiotemporal consistency of social media data and its application in flood models, and provides support for improving the accuracy of flood models. By combining social media data with flood inundation simulation models, the time and location of floods can be predicted more accurately, providing a new technical method for flood warning in data-scarce areas.

[0005] In addition, some studies have built an automated disaster spatiotemporal analysis system by integrating multi-source data such as social media, remote sensing images, and encyclopedias. The system can dynamically track flood events, generate spatiotemporal patterns in real time, and provide important disaster information to decision makers. This dynamic tracking method based on multi-source data effectively makes up for the shortcomings of traditional disaster monitoring systems.

[0006] In actual cases, social media data also demonstrates its important value in post-disaster analysis. Some studies have revealed the spatiotemporal impact patterns of hurricane disasters by analyzing data on social platforms. Studies have shown that spatiotemporal analysis based on social data can not only reveal the specific areas affected by hurricanes, but also provide decision-making basis for post-disaster reconstruction and risk assessment.

[0007] Semantic analysis aims to extract valuable disaster information from unstructured text and provide deeper key information for disaster management. Some studies have combined BERT with TF-IDF and other methods to analyze the topic classification in VGI data and reveal the spatiotemporal evolution and impact patterns of rainstorm events. This study not only identified the key characteristics of the disaster, but also continued to track the development trend of the disaster. Other studies have analyzed the impact of rainstorms and floods through Weibo data, revealing the spatiotemporal consistency between social media activities and rainfall processes, and analyzed the main disaster-stricken areas, indicating that social media data can assess the disaster and provide information support for rapid response.

[0008] Some scholars have developed automated semantic analysis tools for VGI data. For example, based on the architecture of Apache Hadoop and Spark, they process and analyze flood risk information in X data, providing a technical framework for real-time processing of large-scale disaster data. Other studies have built a flood risk assessment framework based on semantic analysis based on the query response model of the ontology, and then used multi-source data to assess the risks in urban floods, successfully developing a disaster management system in a complex urban environment.

[0009] Sentiment analysis provides a new perspective for understanding the psychological impact of disasters on the public. It not only reflects the public's perception and reaction to disasters, but also reveals the public's emotional fluctuations during flood disasters, providing data support for psychological intervention and emergency response strategies. Some studies have conducted sentiment analysis on X tweet data to study the emotional fluctuations of users during natural disasters such as floods. They found that the public's emotional fluctuations during natural disasters are closely related to the severity of the disaster, which helps to understand the public's psychological state and timely disaster crisis management. Another study proposed a framework based on machine learning that can be used to analyze social media big data and model public emotions. It proposed a framework for analyzing social media big data and modeling public emotions, combining machine learning and chart methods to visualize emotional fluctuations on disaster-related topics, providing an intuitive display of emotional trends.

[0010] Sentiment analysis can also be used to evaluate the effectiveness of post-disaster mental health management. Some scholars have revealed the public's emotional fluctuations and their relationship with mental health by analyzing social media data in natural disasters. Studies have found that timely monitoring of public emotional changes can help identify psychological support needs and provide a data basis for post-disaster mental health management. In order to improve the accuracy of sentiment analysis, some studies have used a combination of sentiment dictionaries and machine learning models. For example, the entity sentiment topic model proposed in some studies uses sentiment dictionaries to perform sentiment analysis on disaster-related entities, improves the accuracy and interpretability of tweet sentiment classification, and demonstrates more accurate sentiment evaluation results. In addition, some studies have analyzed disaster response information in social media data, used deep learning models for topic inference and sentiment analysis, demonstrated the relationship between user behavior and climate variables, and provided a new perspective for understanding the public's response to extreme events caused by climate change.

[0011] Exposure analysis is crucial in assessing the risks of population, infrastructure and buildings in flood disasters. By combining data such as population and socioeconomic data, the vulnerability of affected people and assets can be comprehensively assessed. Some studies have used different spatial clustering methods to assess flood exposure in Switzerland. The results show that exposure analysis can effectively identify high-risk areas and provide a basis for the prioritization of flood prevention measures. Other studies have used correlation analysis to assess environmental factors of flood exposure. The study found that environmental factors such as terrain characteristics and rainfall are closely related to flood exposure risks.

[0012] In terms of population exposure research, some studies have used remote sensing images to study the proportion of the global population exposed to floods. The study used MODIS satellite image data to estimate the impact of 913 floods between 2000 and 2018. The results showed that the global population exposed to floods increased by 20-24%. The study also showed that with climate change in the future, the proportion of people exposed to floods will increase further. The study also emphasized the importance of high spatial and temporal resolution observations to better understand the changes in floods and their response measures. In addition, some studies have incorporated flood control facilities into flood exposure analysis, simulating global flood exposure and incorporating the role of dams into flood exposure models for the first time to evaluate the effectiveness of dams in reducing flood-exposed populations under future climate change scenarios. The study found that ignoring the flow regulation function of dams would overestimate flood exposure, while incorporating the regulation function of dams could reduce the flood-exposed population by up to 20.6%.

[0013] Overall, multi-dimensional flood disaster analysis, from spatiotemporal analysis, semantic analysis to sentiment theme analysis and exposure analysis, provides a rich perspective for a comprehensive understanding and management of flood disasters. However, there are still some problems. First, how to effectively integrate the results of multi-dimensional analysis to form a comprehensive disaster analysis framework still needs in-depth research. Second, there is a lack of analysis of disaster types in different regions. Third, the existing population exposure analysis does not combine accurate flooding simulation results to improve the accuracy of the analysis. Summary of the invention

[0014] In response to the problems in the above-mentioned prior art, the present application proposes a multi-dimensional flood disaster analysis method, including disaster spatiotemporal analysis, disaster semantic analysis, response measures adequacy analysis, and population exposure degree analysis; wherein, the disaster spatiotemporal analysis associates the disaster-related information extracted from the text with the time attribute and the spatial coordinate attribute; then the disaster information is analyzed from the perspective of time series and spatial distribution; the time series analysis adopts a quantitative statistical method to analyze the key nodes of flood events from the precipitation trend, flood discharge trend, number of microblogs trend, number of waterlogging points trend, number of help-seeking information trend, and number of measures taken trend; the disaster semantic analysis conducts a thematic analysis on the text of each help-seeking point to mine the semantic information of the disaster in space; the response measures adequacy analysis obtains the adequacy of the disaster response by analyzing the correlation between the number of response measures and the number of help-seeking information; finally, combined with the flood inundation simulation results, the exposure degree of flood disasters to different population groups is evaluated from the dimensions of inundation severity, population distribution, and population structure.

[0015] Optionally, the phenomenon that disasters with similar temporal characteristics have similar spatial distances can be verified from the perspective of the spatial distribution of the help-seeking information; first, the geocoded help-seeking information is projected onto a map, and the kernel density estimation method is used to explore the spatial distribution trend and explore its spatial distribution characteristics.

[0016] Optionally, the spatiotemporal analysis of the disaster situation conducts a topic analysis on the text of each help-seeking point to mine the semantic information of the disaster situation in space; the latent Dirichlet allocation algorithm is used to analyze the theme of the help-seeking information text.

[0017] Optionally, the response measure adequacy analysis includes exploring the data distribution of the assistance request information and the measures taken information, and obtaining the disaster extent of different locations by studying the Spearman correlation between the assistance request information and the measures taken information.

[0018] Optionally, the disaster information extraction uses a fine-tuned ChatGLM-4 model, and the fine-tuned results are evaluated using ROUGE-1 and ROUGE-2 indicators; the ROUGE indicator is used to evaluate the similarity between the content generated by the automatic text generation system and the reference text, and the generation quality is measured by calculating the overlap between the generated text and the reference text at the n-gram and longest common subsequence levels; ROUGE-1 and ROUGE-2 represent the overlap based on a single word and two consecutive words, respectively; the calculation formulas of ROUGE-1 and ROUGE-2 are as follows:

[0019]

[0020] Where Ref represents the reference text, Gen represents the generated text, Count represents the number of times a certain n-gram appears in the text, and min represents the minimum number of times the same n-gram appears in the generated text and the reference text. The numerator in the formula represents the total number of n-grams in the generated text that match the reference text; the denominator represents the total number of all n-grams in the reference text.

[0021] The above technical features can be combined in various suitable ways or replaced by equivalent technical features as long as the purpose of the present invention can be achieved.

[0022] The multi-dimensional flood disaster analysis method provided by the present invention has at least the following beneficial effects compared with the prior art:

[0023] The spatiotemporal evolution characteristics of flood disasters, disaster semantics, adequacy of response measures, and population exposure were analyzed from multiple dimensions. The spatiotemporal analysis revealed the key time nodes and core disaster-affected areas of the disaster, providing a comprehensive understanding of the evolution of the disaster. The semantic analysis used the LDA topic model to mine the main types of disasters in different towns. The analysis of the adequacy of response measures used Spearman correlation analysis to find a strong positive correlation between help-seeking information and response measures (ρ=0.970), and identified areas with insufficient response. The population exposure analysis combined the flood inundation simulation results and WorldPop population data to assess the exposure risks of different regions and age groups. It was found that about 47.92% of the population was in a flood-exposed state, and seven high-risk areas were identified. The above analysis results provide a scientific basis for the formulation of differentiated and precise disaster prevention and mitigation strategies and emergency response measures, which is of great significance for improving regional disaster resilience. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The present invention will be described in more detail below based on embodiments and with reference to the accompanying drawings, wherein:

[0025] Figure 1 A time series analysis graph of the present invention is shown;

[0026] Figure 2 The spatiotemporal heat map of the help-seeking information of the present invention is displayed;

[0027] Figure 3 The spatial distribution characteristic diagram of the assistance information and measures information is shown;

[0028] Figure 4 The disaster situation theme map of each township is displayed;

[0029] Figure 5 Displays the distribution of information on the measures that can be taken for assistance;

[0030] Figure 6 A Spearman correlation analysis plot showing the number of requests for assistance and the number of measures;

[0031] Figure 7 A flood risk classification diagram is shown;

[0032] Figure 8 A map showing the exposure of the population in each township to floods;

[0033] Fig. 9 A map showing the exposure of different age groups of people in each township to floods;

[0034] Fig.10 The ChatGLM-4 model architecture diagram is shown;

[0035] Fig.11 A graph showing the initial extraction results of ChatGLM4 is shown;

[0036] Fig.12 The initial labeled dataset diagram is shown;

[0037] Fig.13 A diagram of the instruction fine-tuning data format is shown;

[0038] Fig.14 The fine-tuned Loss curve and ROUGE indicator graph are shown;

[0039] Fig.15 A graph of the post-optimization results based on ChatGLM-4 is shown. DETAILED DESCRIPTION

[0040] The present invention will be further described below in conjunction with the accompanying drawings.

[0041] The present invention provides a multi-dimensional flood disaster analysis method, including disaster spatiotemporal analysis, disaster semantic analysis, response measure adequacy analysis, and population exposure degree analysis; wherein, the disaster spatiotemporal analysis associates disaster-related information extracted from a text with a time attribute and a spatial coordinate attribute; then the disaster information is analyzed from the perspectives of time series and spatial distribution; the time series analysis adopts a quantitative statistical method to analyze the key nodes of flood events from the precipitation trend, flood discharge trend, microblog number trend, waterlogging point number trend, help-seeking information number trend, and measure-taking number trend; the disaster semantic analysis performs a topic analysis on the text of each help-seeking point to mine the semantic information of the disaster in space; the response measure adequacy analysis obtains the disaster response adequacy by analyzing the correlation between the number of response measures and the number of help-seeking information; finally, combined with the flood inundation simulation results, the exposure degree of flood disasters to different population groups is evaluated from the dimensions of inundation severity, population distribution, and population structure.

[0042] In one embodiment, the spatiotemporal analysis of the disaster situation requires the time attribute and spatial coordinate attribute of the information. The disaster-related information extracted from the text is associated with the time attribute in Doccano. The geographic text is geocoded using the BaiduMap API to obtain the GCJ-02 spatial coordinates, and the GCJ-02 coordinate system is converted to the WGS-84 coordinate system. The data points after the coordinate conversion are superimposed on Google Earth to verify that the error is within 1 meter and will not affect the analysis results. The disaster information is then analyzed from the perspective of time series and spatial distribution. The time series analysis uses quantitative statistics to analyze the key nodes of flood events from the precipitation trend, flood discharge trend, number of Weibo trends, number of waterlogging points trends, number of help messages trends, and number of measures taken trends. The spatial distribution characteristics of the disaster situation were analyzed by using the kernel density estimation (KDE) method to explore the spatial trend of the rescue information. The HDBSCAN (Hierarchical Density-Based Spatial Clustering of Applications with Noise) algorithm was then used to reduce noise. The spatial distribution of the disaster core area was identified by clustering according to spatial distance.

[0043] In one embodiment, time series analysis uses quantitative statistics to combine the time trends of the number of microblogs, the number of help-seeking messages, and the number of measures taken with weather and water conditions information to analyze the key nodes of flood events, including the nodes of their start, peak, and decline stages. The quantitative analysis results are as follows: Figure 1As shown, all charts are statistically analyzed in time series. It should be noted that the flood discharge and precipitation data in the figure are obtained from the Weibo platform, including the Mihe River water level rise and water flow information released by the official Weibo of the hydrological station, and the precipitation information released by the official Weibo of the meteorological station. Figure 1 (a) in the figure shows the rainfall trend in the disaster area. It can be seen that there was heavy rain in the disaster area on the 18th, which was upgraded to heavy rain on the 19th, and turned into moderate rain on the 20th, and then the rain stopped. During this period, there was a light rain on the 27th. The heavy rain on the 19th can be regarded as the main cause of this flood event. Figure 1 (b) in the figure shows the change trend of the flood discharge of the three major reservoirs upstream of the Mi River. According to the flood discharge trend, the heavy rain on the 19th caused the water level of the three major reservoirs upstream to rise rapidly, and the water level quickly approached the safe water level and overflowed the embankment. As can be seen from the figure, the flood discharge was increased on the 19th, and the high-flow flood discharge continued until the 21st, and then the flood discharge was stopped. The Zhiyuan Reservoir began to discharge floods at a small flow on the 23rd, and the light rain on the 27th caused the three major reservoirs to discharge floods on a small scale again, and then maintained a normal flood discharge level. Therefore, the excessive flood discharge caused by heavy rain is the main cause of the flood.

[0044] from Figure 1 In (c), we can find that the number of flood-related microblogs is low from the 18th to the 20th, gradually increases on the 21st, suddenly increases to a peak on the 24th and lasts until the 25th, then suddenly drops on the 26th and then decreases day by day. Based on this, we can judge that the public's attention to the flood is concentrated on the 21st to the 28th. Figure 1 (d) in the figure shows water accumulation information. It can be found that the water accumulation mainly occurs on the 20th to 26th, indicating that the water accumulation gradually recedes after the 26th. Figure 1 (e) in the figure shows the time trend of the number of microblogs containing help-seeking information. The low value area appeared on the 18th and 19th, broke out on the 20th, and remained high until the 27th, with two peaks on the 21st and 25th. Figure 1 (f) in the figure indicates that flood response measures gradually decreased after the 26th and entered the final stage. It can be seen that the problems caused by the flood were mainly concentrated from the 20th to the 26th.

[0045] In order to verify the changes of disaster situation over time, the help-seeking information of each township is listed over time to observe whether the help-seeking information of these townships is concentrated in the same time period, and the disaster timeline is verified accordingly. The township and time information are connected by relationship annotation, such as Figure 2As shown in the figure, the darker the cell color, the more requests for help, and the total number of requests for help from the townships gradually decreases from top to bottom. It can be seen from the figure that the 20th to 28th is the concentrated time period for help information, and it is highly concentrated from the 21st to the 25th. Combined with the above, it can be seen that the timeline of this flood is: on the 19th, the heavy rain reached the safe water level of the reservoir, and the government decided to discharge flood water downstream. On the 20th, excessive flood discharge caused the disaster, which was the starting point of the flood event. From the 21st to the 25th was a severely affected period, during which the disaster further expanded, and the number of various problems caused by the flood increased sharply and remained high. This stage was the peak period of the disaster. The disaster eased on the 26th and 27th, and the 28th was the time point of disaster reduction. Figure 2 It can not only show the change of the degree of disaster over time, but also intuitively show the degree of disaster in each township. As shown in the figure, Shangkou Town, Jitai Town, and Daotian Town are the hardest hit areas. From a global perspective, the disasters in these towns were concentrated from the 21st to the 25th, but the peak periods of disasters in different towns were different. The first law of geography emphasizes spatial interaction. The closer the things are, the higher the degree of correlation. By extension, the changes in disasters are also affected by geographical location. The time series characteristics of disasters in towns that are close should be similar. For example, the time series distribution of the disaster in Shangkou Town is similar to that in Yingli Town and Hou Town, and that in Jitai Town and Daotian Town. The distances between these two groups of towns may be close.

[0046] In one embodiment, the phenomenon that the disaster time series characteristics are similar when the spatial distance is close can be verified from the perspective of the spatial distribution of the help information. First, the geocoded help information is projected onto the map, and the spatial distribution trend is explored using the kernel density estimation method to explore its distribution characteristics in space. The kernel density estimation method represents the density distribution of data in the region by creating a continuous density surface. It is often used to visualize the spatial pattern of point events. The formula is as follows.

[0047]

[0048] f(x) is the estimated density of the help-seeking points at point x, n represents the total number of help-seeking points, xi represents the help-seeking points, h represents the bandwidth, and K represents the kernel function, which is used to calculate the distance weight between a certain position x in the area and the help-seeking point xi. The main parameters of the kernel density estimation method are bandwidth and kernel function. In the study, the distance from the help-seeking data point to the average center of all data points was calculated, and the median distance was 5km, which was used as the kernel density bandwidth. The Gaussian kernel function was selected as the kernel function, which is suitable for density estimation of continuous data and has a good density smoothing effect. The kernel density analysis results are the density distribution of the help-seeking data points in the study area, which provides a reference for subsequent cluster analysis. This study verifies whether the kernel density analysis results are reasonable by observing the kernel density distribution curve and the data histogram distribution.

[0049] The HDBSSCAN algorithm is then used to identify the core distribution of the rescue data points in space. The kernel density estimation method is sensitive to outliers, and when the data density distribution is very different, the fixed bandwidth will cause excessive smoothing of high-density areas and loss of details in low-density areas. The HDBSCAN algorithm does not set global parameters and can identify clusters of different densities. Therefore, the HDBSCAN algorithm has the characteristics of noise reduction and can identify the core distribution of rescue points in space. The parameters of HDBSCAN are the minimum cluster size, and the density trend is obtained according to the kernel density analysis. The cluster size is set to 25. The key steps of the algorithm are shown in Table 1.

[0050] Table 1 HDBSCAN algorithm steps

[0051]

[0052] The HDBSCAN clustering results are visualized in the form of spatial distribution. The rationality of the results is judged by observing the clustering results. Reasonable clustering should make the samples within the class close to each other and the samples between classes separated. Finally, the spatial relationship between the disaster-stricken core areas of Shangkou Town, Yingli Town, Hou Town, Jitai Town and Daotian Town is observed for verification. The results are as follows: Figure 3 This disaster was caused by excessive flood discharge, which resulted in ground waterlogging. The waterlogging point information is also visualized here. Figure 3 From (a) in the figure, we can see that the disaster was concentrated on both sides of the Mi River.

[0053] Figure 3 (b) in the figure shows the spatial distribution and kernel density estimation of the rescue information points. The high-density areas in the figure are distributed along the banks of the Mi River and the Dan River. There is indeed a phenomenon that the disaster time series characteristics are similar when the spatial distance is close. These high-density areas are distributed along the river bank, indicating that the density of rescue points in the study area is quite different. The kernel density estimation is a fixed bandwidth, which may cause excessive smoothing in high-density areas and loss of details in low-density areas. The inherent characteristics of the kernel function make it over- or under-estimate the density in areas with large density changes at the edge of the data. In addition, the kernel density estimation method is sensitive to outliers, and noise will affect the density estimation results. Therefore, after the kernel density estimation completes the spatial trend exploration of the data, the HDBSCAN algorithm is introduced.

[0054] The distribution of dense and sparse areas of data obtained by referring to the kernel density estimation analysis as prior knowledge helps the HDBSCAN algorithm to more accurately identify and divide clusters of different densities. When the cluster size is set to 25, the HDBSCAN cluster distribution is consistent with the kernel density estimation distribution, so 25 is used as the cluster parameter. Setting the cluster size too large may cause small but important clusters to be ignored, while setting it too small may produce too many noise clusters, affecting the interpretability of the results. Figure 3As shown in (c) in the figure, the HDBSCAN algorithm clusters the help-seeking points into 6 clusters, indicating that there are 6 density cores. In the kernel density estimation results, there are high-density areas in Houzhen and Shengcheng Street, but because there are not many help-seeking information points, they are classified as noise points by the HDBSCAN algorithm. There is a high correlation between help-seeking and response measures. The minimum cluster is also set to 25 to ensure consistent scale and cluster the measure information. Figure 3 As shown in (d) in the figure, we can find that the clustering results of response measures include 5 clusters, and the spatial distribution of the clusters is basically consistent with the clustering results of assistance requests.

[0055] In one embodiment, the temporal distribution and spatial distribution of the disaster situation are studied to explain the disaster situation from the perspective of time and space. Affected by topography, surface cover and social organization factors, the disaster situation in different regions is not the same. The text of each help point can be subject analyzed to mine the semantic information of the disaster situation in space. The study uses the Latent Dirichlet Allocation (LDA) algorithm to analyze the topic of the help information text. After the help text is processed by word segmentation and stop word removal, the key steps of the LDA algorithm are shown in Table 2.

[0056] Table 2 - LDA algorithm steps

[0057]

[0058] The help information contains the location and disaster text. The disaster themes of different regions can be analyzed from the perspective of text themes, so as to mine the disaster manifestations of each township. The help information text is summarized according to townships, and the Hanlp tool is used to segment the text, remove stop words and construct a bag of words. The LDA algorithm is then used to analyze the text theme. The theme is composed of several keywords and their corresponding probabilities. The three keywords with the highest probability in each township are taken out, as shown in Table 3.

[0059] Table 3 Keyword composition of disaster themes in each township

[0060]

[0061] Visualize the disaster themes of each township and study the thematic distribution characteristics of each township. Figure 4As shown in the figure, the disaster themes of towns and villages show a clustering phenomenon in space. For example, Shangkou Town is adjacent to Hou Town and both suffer from material shortages. Gucheng Street is adjacent to Shengcheng Street and both suffer from traffic problems. This shows that the closer the regions are, the higher the degree of correlation of disasters. This phenomenon is in line with the first law of geography, that is, "all things are interrelated, but things near are more closely related than things far away." Overall, the distribution of disaster themes between towns and villages is not completely related to spatial distance. This phenomenon may be related to the topography, socioeconomic conditions and infrastructure conditions in the region. For example, towns and villages far away from urban areas are prone to material shortages, and towns and villages downstream of rivers are more likely to have drainage problems. This is also in line with the second law of geography, that is, the "law of spatial heterogeneity." Disaster emergency management needs to take into account the specific conditions and needs of each region. Analyzing the spatial themes of disasters is of great significance and can provide important reference for emergency response and resource allocation.

[0062] In order to further improve the accuracy of information extraction, the study introduced the ChatGLM-4 9B Instruct (hereinafter referred to as ChatGLM-4) model to perform post-optimization processing on the extracted information, and used its context understanding and reasoning advantages to enhance the accuracy and quality of information extraction. At the same time, the extracted information also provides output constraints for the large model, thereby reducing the hallucination problem of the ChatGLM-4 model. Fig.10 The model architecture of ChatGLM-4 is based on the Transformer framework. The core structure includes input layer, word embedding layer, multi-layer GLMBlock and root mean square normalization layer (RMSNorm). It also integrates technologies such as Residual Network (ResNet) and Rotary Position Embedding (RoPE).

[0063] The input layer first converts the text sequence into a sequence of tokens through Tokenizer, and then converts these tokens into unique numeric identifiers through Input_ids. These identifiers are then passed to the word embedding layer, where each identifier is mapped to a high-dimensional vector space to generate the initial input representation of the model. The core of ChatGLM-4 is a deep structure consisting of 40 GLMBlocks. Each GLMBlock contains a self-attention mechanism and a feed-forward neural network (FFN) component. The self-attention mechanism first maps the input to the query (Q), key (K), and value (V) matrices through a linear transformation. Then, the rotation position encoding is introduced to enhance the model's perception of sequence position information. Next, the attention calculation is performed, and finally the result is converted back to the original dimension through a linear transformation. The feed-forward neural network part contains two linear transformation layers, with the SwiGLU activation function used in the middle, which can provide better nonlinear transformation capabilities than the traditional ReLU activation function. The intermediate layer dimension of the feed-forward neural network is expanded to 13696, further enhancing the representation ability of the model.

[0064] After each self-attention and feedforward neural network component, ChatGLM-4 applies residual connections, Dropout regularization, and RMS normalization. Residual connections enable information to be directly transferred from shallower layers to deeper layers, effectively improving the gradient vanishing problem and promoting gradient flow in deep networks. Dropout is a regularized network that prevents overfitting by randomly dropping neurons during training. RMS normalization replaces traditional layer normalization (LayerNorm) and stabilizes the training process by standardizing the activation values ​​of neurons, accelerating convergence and improving the generalization ability of the model.

[0065] After the GLMBlock stack, the model uses an additional RMS normalization layer for final normalization. The output layer converts the high-dimensional representation back to the vocabulary space and finally goes through a decoding process to generate human-readable text output.

[0066] The innovation of ChatGLM-4 lies in the use of Grouped Query Attention (GQA) mechanism and the use of SwiGLU activation function instead of traditional ReLU. Grouped query attention reduces computational complexity and memory usage by reducing the number of attention heads while still maintaining model performance. The SwiGLU activation function has stronger nonlinear transformation capabilities, which helps capture more complex features and patterns. In short, the architectural design of the ChatGLM-4 model achieves a balance between efficiency and performance, enhancing computational efficiency and model generalization capabilities while maintaining language understanding and generation capabilities.

[0067] For the specific task of flood disaster information extraction and analysis, directly using the general ChatGLM-4 model cannot output standardized information results, such as Fig.11 shown. To standardize the model output and enhance the model's ability to extract and analyze flood disaster information, the study fine-tunes the ChatGLM-4 model using the Low-Rank Adaptation (LoRA) technique.

[0068] 1. Fine-tuning principle. LoRA is a parameter-efficient fine-tuning method that can adapt to downstream tasks by introducing a small number of trainable parameters while keeping the original model parameters unchanged. The core idea of LoRA is to add a low-rank update matrix ΔW to the weight matrix W0 of the pre-trained model, i.e., W = W0 + ΔW. ΔW is represented as the product of two low-rank matrices: ΔW = BA, where B ∈ R^(d×r) and A ∈ R^(r×k), and r << min(d, k). During training, W0 remains frozen, and only A and B contain trainable parameters. This method can significantly reduce the number of parameters to be trained while retaining most of the model's capabilities.

[0069] When fine-tuning the ChatGLM-4 model, LoRA is mainly applied to the query (Q), key (K), and value (V) matrices in the self-attention mechanism, as well as the weight matrices of the feed-forward neural network. For each weight matrix that needs to be adapted, LoRA introduces a pair of low-rank matrices A and B. During the forward propagation process, the model's calculation becomes h = W0x + BAx. Here, x is the input vector, and h is the output vector. By adjusting matrices A and B, the model can learn the fine-tuning required for specific tasks.

[0070] Compared with full-parameter fine-tuning, the advantages of LoRA are as follows: First, high parameter efficiency. Compared with full-parameter fine-tuning, LoRA only needs to train a small number of parameters, significantly reducing the computational and storage requirements. Second, strong flexibility. Different LoRA modules can be trained for different tasks, and tasks can be quickly switched without storing the complete fine-tuned model. Third, no inference latency is introduced. During inference, ΔW can be merged with W0 without increasing additional computational overhead. Fourth, the initial model capabilities are maintained. Since the original weights remain unchanged, the model does not lose the knowledge learned during the pre-training stage. By fine-tuning ChatGLM-4 using the LoRA technique, the model can better adapt to the flood disaster information extraction and analysis task, output more standardized and task-related information, and at the same time maintain the model's general capabilities in other fields.

[0071] 2. Fine-tuning data preparation. Fine-tuning data and annotated data from named entity recognition, research on converting the jsonl format annotation file exported by the Doccano annotation platform into a question-answering format that satisfies instruction fine-tuning. Annotated dataset examples are as follows: Fig.12 shown.

[0072] Then write code to convert the labeled dataset into the instruction question answering format. The converted sample is as follows: Fig.13 shown.

[0073] 3. Model fine-tuning. The fine-tuning environment is built based on AutoDL (www.autodl.com), with a Linux system, L40 graphics card (48GB), AMD EPYC 7T83 processor, and 100GB of running memory. Based on comprehensive consideration of model characteristics, task requirements, and computing resources, the final core parameter settings are as follows:

[0074] Maximum input length (max_input_length) and maximum output length (max_output_length): Both are set to 512 tokens. The length of 512 can match the length of most microblog texts while keeping the computational complexity low.

[0075] The maximum number of training steps (max_steps) is set to 3000. The learning rate (learning_rate) is set to 1e-4, a learning rate. The per-device training batch size (per_device_train_batch_size) is set to 1, which takes into account the size of the model and the available GPU memory. The gradient accumulation steps (gradient_accumulation_steps) is set to 2 steps, which allows the model to accumulate multiple batches of gradients before performing a parameter update.

[0076] The save_steps setting is set to save the model every 300 steps. The per-device evaluation batch size is set to 1, which is consistent with the training batch size, ensuring that the evaluation process uses the same computing resource configuration as the training process. The evaluation steps (eval_steps) are also set to 300 steps, which is the same as the save steps.

[0077] The results after fine-tuning are evaluated using ROUGE-1 and ROUGE-2 indicators. The ROUGE (Recall-Oriented Understudy for Gisting Evaluation) indicator is used to evaluate the similarity between the content generated by the automatic text generation system and the reference text. It mainly measures the generation quality by calculating the overlap between the generated text and the reference text at the level of n-gram, longest common subsequence, etc. ROUGE-1 and ROUGE-2 represent the degree of overlap based on 1-gram (single word) and 2-gram (two consecutive words), respectively. The calculation formulas of ROUGE-1 and ROUGE-2 are shown below.

[0078]

[0079] Where Ref represents the reference text, Gen represents the generated text, Count represents the number of times a certain n-gram appears in the text, and min represents the minimum number of times the same n-gram appears in the generated text and the reference text. The numerator in the formula represents the total number of n-grams in the generated text that match the reference text. The denominator represents the total number of all n-grams in the reference text.

[0080] The Loss curve and ROUGE index of the fine-tuning process are as follows: Fig.14 As shown. After 1500 steps of training, a total of 1.46 epochs, the model's loss function (Loss) gradually decreased from the initial 0.7888 to 0.16, indicating that the model gradually converged on the training set. ROUGE-1 increased from the initial 76.31 to the final 83.19, an increase of 9%. This shows that the model's recall rate at the word level has been significantly improved, and its ability to capture keywords has been improved. ROUGE-2 increased from 58.74 to 70.02, an increase of 19%, indicating that the model has made greater progress in its ability to capture two consecutive words and has been improved in maintaining the semantic coherence of the original text.

[0081] The ROUGE-1 and ROUGE-2 indicators tended to stabilize in the later stages of training, fluctuating around 83 and 70 respectively, indicating that the ROUGE-1 and ROUGE-2 indicators of the model have basically converged. The entire training process took 1 hour and 31 minutes, and the fine-tuned model was significantly improved compared to the initial state, indicating that the fine-tuning process improved the model's ability to extract and analyze flood disaster information. Finally, the fine-tuned model increment was merged with the initial model to form a complete small domain model for flood information extraction.

[0082] The core idea of ​​post-optimization processing is to guide the ChatGLM-4 model to compare and optimize the initial extraction results by constructing prompt statements. The key to the prompt statement is to input the preliminary results obtained based on BERT-Bi-LSTM-CRF and rule extraction methods into the model, and use its language understanding ability to compare and analyze the preliminary extraction results with the original content, identify possible errors or omissions, and provide more accurate and comprehensive extraction results. The key elements of the sentence prompt constructed in the study are as follows:

[0083] Initial text: Provide the complete disaster report text to provide comprehensive contextual information for the model.

[0084] Task description: Clearly state that information related to flood disasters needs to be extracted from the text, and list the specific categories that need to be extracted, such as time, location, weather, current issues, measures taken, water conditions, water bodies, organizations, losses, needs, people and contact information, etc.

[0085] Preliminary extraction results: Show the model the preliminary results based on BERT-Bi-LSTM-CRF and rule extraction methods to provide a reference for subsequent optimization of the model.

[0086] Optimization instructions: Require the model to give revised information extraction results based on the preliminary extraction results.

[0087] The output of the optimized model is as follows: Fig.15 As shown in the figure, the yellow dialogue is the designed prompt statement, and the gray-white is the model output result.

[0088] As can be seen from the figure, the extraction result of the traditional method classifies the first sentence "urgently need high-power water pumps, bedding, mineral water and other materials" as a current problem, but it should actually be classified as a demand like the second sentence. The fine-tuned ChatGLM-4 model combines the prompt statement to optimize the extraction of information and improve the information extraction quality of the text. The prompt command is used to make ChatGLM-4 output the post-optimization information as a table, and the script is used to compare the annotated true value and the information extracted after optimization. The post-optimization accuracy is shown in Table 4. It can be found that the overall Precision of the information extraction quality optimization scheme is 92.55%, Recall is 93.60%, and F1 is 92.51%, which has greatly improved the accuracy of flood information extraction.

[0089] Table 4 Information post-optimization extraction accuracy based on ChatGLM-4 model

[0090]

[0091] In one embodiment, in the study of flood disasters, analyzing the relationship between response measures and the number of help requests can reveal the adequacy of the response measures and the effectiveness of resource allocation. If the number of response measures is highly positively correlated with the number of help requests, this indicates that the disaster response is adequate; if the correlation is low, it may indicate that resource allocation does not meet actual needs. The analysis of the adequacy of measures helps optimize future disaster response strategies and ensure the accurate allocation of resources. It should be noted that flood disasters are affected by many factors, and disaster information in each affected township does not necessarily follow a normal distribution. Therefore, it is necessary to observe the data distribution before choosing to use the correlation analysis method.

[0092] First, explore the data distribution of the rescue information and the measures taken, such as Figure 5 As shown. The frequency distribution of the data of the assistance information and the measures information does not meet the normal distribution assumption, so the study uses Spearman correlation analysis. Spearman correlation analysis is a non-parametric statistical method used to measure the monotonic relationship between two variables. Compared with Pearson correlation, Spearman correlation does not require a linear relationship between variables and is suitable for situations where the data distribution does not meet the normal distribution or there are outliers. The Spearman formula is shown below.

[0093]

[0094] Where ρ represents the Spearman correlation coefficient, d i is the difference between the i-th pair of rankings, and n represents the number of samples. The value range of the Spearman correlation coefficient is [-1, 1]. When ρ>0, it means that there is a positive correlation between the variables; when ρ<0, it means that there is a negative correlation between the variables; when ρ=0, it means that there is no obvious correlation between the variables.

[0095] The results of Spearman correlation analysis are as follows Figure 6 As shown, each scatter point represents a town, with different colors and distinctions. From the figure, we can see an obvious positive correlation trend, which means that when the number of help-seeking information increases, the corresponding number of response measures also increases. The Spearman rank correlation coefficient is 0.970, indicating that there is a very strong positive correlation between the number of help-seeking information and the number of measures. In addition, the P value is 0.0000, indicating that this correlation is statistically significant, that is, the correlation is unlikely to be caused by random variation.

[0096] Based on the results of the Spearman correlation coefficient, we can infer that in this data set, there is a strong positive correlation between the number of disaster response measures and the number of help-seeking messages. Combined with the results of the spatial characteristic analysis, among the five towns that were most severely affected, namely Yangkou Town, Yingli Town, Shangkou Town, Jitai Town and Daotian Town, Jitai Town had insufficient response measures to floods, while other towns had relatively sufficient response measures.

[0097] In one embodiment, spatiotemporal analysis, semantic analysis, and correlation analysis comprehensively explore the evolution law of flood disasters in the study area, spatial distribution characteristics, and the relationship between various types of rescue information and response measures. These analyses provide an important basis for understanding the overall situation of the disaster and its impact on different regions. The ultimate impact of flood disasters is reflected in the threat to human life and property safety, and the analysis of population flood exposure can intuitively reflect the differences in flood risks faced by people in different regions, which helps to more accurately assess the potential social impact of disasters and provide important support for the formulation of targeted emergency response and evacuation strategies. The study will combine the flood inundation simulation results and population distribution data in the previous article to analyze the flood exposure of different age and gender groups in the study area. The population distribution, age and gender structure data are derived from the WorldPop population dataset in 2018, with a spatial resolution of 100 meters. Combined with the flood inundation simulation results, the exposure of different population groups to flood disasters is evaluated from the dimensions of inundation severity, population distribution, and population structure (age, gender).

[0098] The exposure risk of people varies at different flooding depths. This study set a classification standard for flooding severity based on the following key water depths: 15cm, 30cm, 40cm and 60cm.

[0099] 1. The boundary between slight waterlogging and waterlogging (15cm). When the water depth does not exceed 15cm, it can be considered as slight waterlogging. Generally, the height of the curb of a city sidewalk is between 15-20cm. Waterlogging below this height usually does not significantly affect the normal passage of pedestrians and vehicles. Waterlogging exceeding this depth is defined as waterlogging, and it is difficult to distinguish the underwater ground, which affects the travel of people and vehicles.

[0100] 2. The boundary between low risk and medium risk (30cm): When the water depth exceeds 30cm, the exhaust pipes of some models may be submerged, and the engine may be turned off, resulting in road traffic closure. Considering that the exhaust pipe of an ordinary car is 20-30cm above the ground, while that of an SUV is about 30-40cm, 30cm is set as the boundary between low risk and medium risk.

[0101] 3. Medium-risk and high-risk boundary (40cm): When the water depth reaches 40cm, it may cause water to enter the ground floor buildings. The building site is generally about 20cm higher than the road, and the entrance steps are usually no more than 15cm, plus the curb height, the total is about 40cm. Therefore, water accumulation exceeding this depth may cause direct damage to the building.

[0102] 4. The boundary between high risk and extremely high risk (60cm): When the water depth exceeds 60cm, it may pose a serious threat to personal safety. At this depth, it is difficult for adults to stand up. Once they fall, their respiratory tract may be below the water surface, causing life-threatening danger. At the same time, this depth will also cause water to enter the air intake system of most vehicles, causing traffic paralysis. Based on the above analysis, this study divides flood inundation risks into five levels, as shown in Table 4.

[0103] Table 4 Classification of flooding risk

[0104]

[0105] After completing the flooding risk classification, the maximum flooding depth simulation results are divided into five corresponding levels according to the water depth. The flooding risk classification after classification is as follows: Figure 7 shown.

[0106] The study uses the 100-meter resolution population distribution dataset provided by the WorldPop project as the basic data source for population spatial distribution. The WorldPop data format is GeoTIFF, including total population density and population density layers divided by 5-year age groups and gender. In 2018, there were about 1.1031 million people in S City. According to the risk zone classification data, the total population exposure status in each risk zone was statistically analyzed. It can be seen that 528,600 people were exposed to floods (water depth of more than 15 cm) in this flood disaster, indicating that 47.92% of the population lives in areas susceptible to floods. The exposed population in different towns was then analyzed, such as Figure 8 shown.

[0107] Figure 8 The figure describes the proportion of flood inundation areas, the number of affected people, and the exposure intensity (exposed population density) of 15 towns in S City. The figure uses the percentage of flooded area in each town as the x-axis and the number of exposed population as the y-axis. The size of the bubble indicates the size of the exposed population, and the color indicates the ratio of the exposed population density to the flooded area.

[0108] Under low exposure conditions, Shengcheng Street has the largest number of exposed people and a high density of exposed people. Luocheng Street is second, while Hualong Town and Hou Town have more exposed people, but the exposure intensity is not as high as Tianliu Town and Wenjia Street. Under moderate exposure conditions, Shengcheng Street is still the area with the largest exposed population, Gucheng Street and Luocheng Street also have a high number of exposed people, but the density varies. Although the number of exposed people in Tianliu Town and Wenjia Street is not as high as the above three areas, their exposed population density is high.

[0109] Under high exposure conditions, Shengcheng Street is still the area with the largest number of exposed people, and the exposure intensity is also at the highest level, followed by Tianliu Town and Luocheng Street, which also show a high number and density of exposed population. Yingli Town, Hou Town, Gucheng Street, and Shangkou Town began to show a higher exposure risk, and their exposed population and density are relatively high. Under extremely high exposure conditions, Shengcheng Street is not only at the highest level in terms of both the number of exposed people and the intensity of exposure, followed by Tianliu Town. Shangkou Town was not significant in the previous exposure level, but the risk increased significantly under the extremely high exposure level. On the whole, Shengcheng Street, Shangkou Town, Tianliu Town, Yingli Town, Jitai Town, Luocheng Street, and Daotian Town can be regarded as the highest risk areas for exposure.

[0110] People of different age groups have different abilities to cope with flood risks. The study used the WorldPop data using a raster calculator to reclassify age groups into four categories: 0-14 years old (children), 15-19 years old (youth), 20-64 years old (adults), and 65 years old and above (elderly). Children are more vulnerable and often lack the ability to protect themselves during flood events, requiring special attention. Young people are physically stronger, but due to lack of experience, they may still face higher risks. Adults are relatively mature mentally and have better coping skills in the face of floods. Elderly people are generally considered to be at high risk due to their physical strength, health and mobility limitations.

[0111] according to Fig. 9 It can be seen that Shengcheng Street shows the highest risk in all exposure levels. At the extremely high exposure level, its adult population accounts for the largest proportion, but the absolute number of children and the elderly population is also relatively high, and the overall vulnerability is high. Luocheng Street has a higher risk in the low and medium exposure levels, but the relative risk decreases in the high and extremely high exposure levels. Its age structure is relatively balanced, and its overall vulnerability is lower than that of Shengcheng Street. In the moderate exposure area of ​​Gucheng Street, the exposure ratio of children is relatively large. Due to the vulnerability of children in floods, the township needs to formulate flood prevention and safety measures for children to increase the safety of children in flood events. In the high and extremely high exposure areas, the exposure ratio of children and young people in Tianliu Town is relatively high, and attention should also be paid to the evacuation and resettlement measures of young people. The exposure risk of Shangkou Town and Yingli Town increases sharply at the high and extremely high exposure levels, among which the proportion of children and the elderly population is relatively high. It is necessary to strengthen flood prevention and protection measures for children and the elderly to ensure the targeted rescue.

[0112] Although the present invention is described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the present invention. It should therefore be understood that many modifications may be made to the exemplary embodiments and that other arrangements may be devised without departing from the spirit and scope of the present invention as defined by the appended claims. It should be understood that the various dependent claims and features described herein may be combined in a manner different from that described in the original claims. It should also be understood that features described in conjunction with individual embodiments may be used in other described embodiments.

Claims

1. A multi-dimensional flood disaster analysis method, characterized in that: It includes spatiotemporal analysis of disaster situation, semantic analysis of disaster situation, analysis of adequacy of response measures, and analysis of population exposure. Among them, the spatiotemporal analysis of disaster situation associates the disaster-related information extracted from the text with the time attribute and the spatial coordinate attribute. Then, the disaster information is analyzed from the perspective of time series and spatial distribution. The time series analysis uses quantitative statistics to analyze the key nodes of flood events from the trends of precipitation, flood discharge, number of microblogs, number of waterlogging points, number of help-seeking messages, and number of measures taken. The semantic analysis of disaster situation conducts thematic analysis on the text of each help-seeking point to mine the semantic information of the disaster situation in space. The analysis of adequacy of response measures derives the adequacy of disaster response by analyzing the correlation between the number of response measures and the number of help-seeking messages. Finally, combined with the results of flood inundation simulation, the exposure of different population groups to flood disasters is evaluated from the dimensions of inundation severity, population distribution, and population structure.

2. The multi-dimensional flood disaster analysis method according to claim 1 is characterized in that: The phenomenon that disasters with similar temporal characteristics have similar spatial distances is verified from the perspective of the spatial distribution of the help-seeking information. The geocoded help-seeking information is first projected onto the map, and the kernel density estimation method is used to explore the spatial distribution trend and explore its spatial distribution characteristics.

3. The multi-dimensional flood disaster analysis method according to claim 1 is characterized in that: The spatiotemporal analysis of the disaster situation conducts a thematic analysis on the texts of each help-seeking point to mine the semantic information of the disaster situation in space; the latent Dirichlet allocation algorithm is used to analyze the theme of the help-seeking information text.

4. The multi-dimensional flood disaster analysis method according to claim 1 is characterized in that: The analysis of the adequacy of response measures includes exploring the data distribution of assistance information and measures taken information. By studying the Spearman correlation between assistance information and measures taken information, the degree of disaster damage in different locations is obtained.

5. The multi-dimensional flood disaster analysis method according to claim 1 is characterized in that: The fine-tuned ChatGLM-4 model was used to extract disaster information. The fine-tuned results were evaluated using the ROUGE-1 and ROUGE-2 indicators. The ROUGE indicator is used to evaluate the similarity between the content generated by the automatic text generation system and the reference text. The generation quality is measured by calculating the overlap between the generated text and the reference text at the n-gram and longest common subsequence levels. ROUGE-1 and ROUGE-2 represent the overlap between a single word and two consecutive words, respectively. The calculation formulas for ROUGE-1 and ROUGE-2 are as follows: Where Ref represents the reference text, Gen represents the generated text, Count represents the number of times a certain n-gram appears in the text, and min represents the minimum number of times the same n-gram appears in the generated text and the reference text. The numerator in the formula represents the total number of n-grams in the generated text that match the reference text; the denominator represents the total number of all n-grams in the reference text.

Images