Efficient flood resilience assessment method and system based on natural language processing and data-driven model

Abstract

The present invention discloses an efficient flood resilience assessment method and system based on natural language processing and data-driven models, including collecting multi-source data of the study area, constructing a spatiotemporal data cube, performing data quality assessment, and obtaining a spatiotemporal data cube; constructing a text corpus, performing multimodal text preprocessing, constructing a knowledge graph of flood resilience, and generating a preliminary resilience index system; classifying resilience indicators to obtain an index set; extracting spatiotemporal sequence data, performing correlation and causality analysis, and obtaining a final optimized resilience index system; performing feature engineering to generate a feature matrix; adding semantic features to the feature matrix; calling a quantitative method to calculate a target variable; constructing a data set including a feature matrix and a target variable, constructing a resilience assessment model, and generating a resilience assessment report. The present invention realizes the dynamic adjustment of the index system, improves the computational efficiency of the model, and enhances the expression ability and prediction accuracy of the model.

Description

Technical Field

[0001] The present invention relates to disaster simulation and assessment technology, in particular to an efficient flood resilience assessment method and system based on natural language processing and data-driven models. Background Art

[0002] With the acceleration of global climate change and urbanization, the threat of flood disasters to urban systems is becoming increasingly severe. Urban flood resilience, as an important component of urban resilience, has become a key indicator to measure the city's ability to resist, recover and adapt to flood disasters. Improving urban flood resilience has become a key issue in urban sustainable development and disaster management. As an important tool for understanding and improving urban flood resilience, flood resilience assessment is of great significance for formulating effective disaster prevention and mitigation strategies, optimizing resource allocation and improving emergency response capabilities. Accurate, comprehensive and dynamic resilience assessment can not only help urban managers identify vulnerable links in the system, but also provide a scientific basis for formulating targeted resilience improvement measures. In addition, efficient assessment methods can support real-time monitoring and rapid decision-making, which is particularly critical when responding to sudden flood events. Therefore, developing a flood resilience assessment method that can integrate multi-source information, adapt to complex urban systems, and respond quickly has great practical significance and social value for improving the overall flood resistance of cities and reducing disaster losses.

[0003] To this end, technicians have conducted a lot of research, but how to effectively integrate these technologies to achieve efficient, accurate and refined assessments remains a challenge. The insufficient level of refinement in resilience assessments results in coarse spatiotemporal granularity in resilience assessment results, making it difficult to guide the precise prevention and control of urban flood disasters. Existing indicator systems are often too static and rigid to adapt to the dynamic changes in urban systems and the particularities of different cities. How to build a comprehensive and flexible indicator system that can be dynamically adjusted to reflect real-time changes in urban resilience is an urgent problem to be solved. In addition, when processing large-scale, high-dimensional spatiotemporal data, the computational efficiency and prediction accuracy of existing models still need to be improved, especially when facing extreme events and long-term trend forecasts. The generalization ability and robustness of the model are insufficient, making it difficult to achieve precise prevention and control of urban flood disasters. Therefore, research and innovation are needed. Summary of the invention

[0004] The purpose of the invention is to provide an efficient flood resilience assessment method and system based on natural language processing and data-driven models to solve the above-mentioned problems existing in the prior art.

[0005] The technical solution is an efficient flood resilience assessment method based on natural language processing and data-driven models, which includes the following steps:

[0006] Step S1, collect multi-source data of the study area, and perform data cleaning, standardization and spatiotemporal alignment on the multi-source data to construct a spatiotemporal data cube; perform data quality assessment on the spatiotemporal data cube; obtain and store a spatiotemporal data cube that meets the expected quality, wherein the multi-source data includes meteorological, hydrological, remote sensing, population, economic and drainage system data;

[0007] Step S2: extract geographic identifiers of the study area from the spatiotemporal data cube and construct a text corpus; perform multimodal text preprocessing on the text corpus, and use natural language processing technology to perform keyword extraction, dynamic topic modeling, entity recognition, and relationship extraction to obtain basic information of the graph; construct a knowledge graph of flood resilience based on the basic information of the graph, and use the knowledge graph to generate a preliminary resilience indicator system;

[0008] Step S3: Based on the knowledge graph, classify the resilience indicators in the preliminary resilience indicator system, obtain the indicator set, and perform independence test and weight calculation; use the spatiotemporal data cube to extract spatiotemporal series data for the indicators, and perform correlation and causality analysis; use the indicator quantification method and use the historical data in the spatiotemporal data cube to calculate the dynamic benchmark value; obtain the final optimized resilience indicator system; the resilience indicator system includes indicator definition, classification, weight and quantification method;

[0009] Step S4: Based on the resilience indicator system and the spatiotemporal data cube, perform feature engineering to generate a feature matrix; combine the knowledge graph to add semantic features to the feature matrix; call the quantitative method to calculate the target variable; construct a data set containing the feature matrix and the target variable, and divide it into a training set and a test set; build a resilience assessment model and train it through the training set; when in use, call the trained resilience assessment model to generate a resilience assessment report.

[0010] According to another aspect of the present application, there is also provided an efficient flood resilience assessment system based on natural language processing and data-driven models, comprising:

[0011] at least one processor; and,

[0012] a memory communicatively connected to at least one of the processors; wherein,

[0013] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the efficient flood resilience assessment method based on natural language processing and data-driven models as described in any of the above technical solutions.

[0014] Beneficial effects: The present invention realizes dynamic adjustment of the index system by constructing a resilience index system, overcoming the static and rigid problems of the traditional index system; generates a feature matrix through feature engineering, and adds semantic features in combination with the knowledge graph, thereby improving the computational efficiency of the model, and enhancing its adaptability to complex urban systems, and enhancing the model's expression ability and prediction accuracy; through accurate, comprehensive and dynamic resilience assessment, it optimizes resource allocation, improves emergency response capabilities, and enhances the city's overall flood resistance. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flow chart of the present invention.

[0016] Figure 2 This is a flow chart of step S1 of the present invention.

[0017] Figure 3 This is a flow chart of step S2 of the present invention.

[0018] Figure 4 This is a flow chart of step S3 of the present invention.

[0019] Figure 5 This is a flow chart of step S4 of the present invention. DETAILED DESCRIPTION

[0020] like Figure 1 As shown, this application proposes an efficient flood resilience assessment method based on natural language processing and data-driven models, including the following steps:

[0021] Step S1, collect multi-source data of the study area, and perform data cleaning, standardization and spatiotemporal alignment on the multi-source data to construct a spatiotemporal data cube; perform data quality assessment on the spatiotemporal data cube; obtain and store a spatiotemporal data cube that meets the expected quality, wherein the multi-source data includes meteorological, hydrological, remote sensing, population, economic and drainage system data;

[0022] Step S2: extract geographic identifiers of the study area from the spatiotemporal data cube and construct a text corpus; perform multimodal text preprocessing on the text corpus, and use natural language processing technology to perform keyword extraction, dynamic topic modeling, entity recognition, and relationship extraction to obtain basic information of the graph; construct a knowledge graph of flood resilience based on the basic information of the graph, and use the knowledge graph to generate a preliminary resilience indicator system;

[0023] Step S3: Based on the knowledge graph, classify the resilience indicators in the preliminary resilience indicator system, obtain the indicator set, and perform independence test and weight calculation; use the spatiotemporal data cube to extract spatiotemporal series data for the indicators, and perform correlation and causality analysis; use the indicator quantification method and use the historical data in the spatiotemporal data cube to calculate the dynamic benchmark value; obtain the final optimized resilience indicator system; the resilience indicator system includes indicator definition, classification, weight and quantification method;

[0024] Step S4: Based on the resilience indicator system and the spatiotemporal data cube, perform feature engineering to generate a feature matrix; combine the knowledge graph to add semantic features to the feature matrix; call the quantitative method to calculate the target variable; construct a data set containing the feature matrix and the target variable, and divide it into a training set and a test set; build a resilience assessment model and train it through the training set; when in use, call the trained resilience assessment model to generate a resilience assessment report.

[0025] This embodiment builds a comprehensive and accurate data foundation for flood resilience assessment through the collection, preprocessing and fusion of multi-source heterogeneous data, thereby improving the reliability and accuracy of subsequent assessments. Specifically, by introducing multimodal data fusion technologies, such as tensor decomposition and deep learning methods (such as multimodal autoencoders), seamless integration of multi-dimensional data such as meteorology, hydrology, remote sensing, and socio-economics is achieved, effectively capturing the multi-faceted characteristics of flood resilience. In particular, when processing high-dimensional spatiotemporal data, advanced spatiotemporal interpolation algorithms (such as spatiotemporal kriging) and dynamic time warping (DTW) technology are used to improve the spatiotemporal consistency and continuity of data, providing high-quality input for subsequent resilience assessments. In addition, by introducing deep learning-based anomaly detection algorithms (such as variational autoencoders VAE) and multi-scale data quality assessment indicators, data anomalies can be automatically identified and processed, and data quality can be accurately quantified, thereby improving data reliability. In terms of data storage and indexing, distributed spatiotemporal databases (such as Apache Sedona) and multidimensional index structures (such as mixed indexes of R-trees and quadtrees) are used to achieve efficient storage and rapid retrieval of large-scale spatiotemporal data, laying the foundation for real-time flood resilience assessment. It not only improves the efficiency of data processing (processing speed increased by about 50%), but also achieves accurate quantification of data quality (accuracy increased by about 30%), providing more comprehensive, accurate and reliable data support for subsequent flood resilience assessments, thereby enhancing the scientificity and credibility of the assessment results.

[0026] By combining natural language processing technology with flood resilience assessment, a breakthrough has been achieved in automatically extracting resilience-related knowledge and indicators from unstructured text data. A natural language processing tool chain specifically for the field of flood resilience has been developed, including domain-specific pre-trained language models (such as HydroLM based on BERT) and a multimodal information extraction framework. In particular, in the text preprocessing stage, a document structure understanding algorithm based on graph neural networks was introduced, which can effectively capture the hierarchical relationships and logical structures in the document and improve the accuracy of subsequent knowledge extraction. In terms of keyword extraction and topic modeling, the attention mechanism and dynamic topic model (such as Dynamic LDA) are combined, which can not only identify static important concepts, but also track the dynamic evolution of resilience-related topics, providing important support for understanding the long-term trend of flood resilience. Methods based on remote supervision and transfer learning are introduced in the entity recognition and relationship extraction links, which effectively solve the problem of insufficient annotated data of domain expertise and improve the performance of the model in identifying professional terms and complex relationships (the F1 score is increased by about 25%). In the knowledge graph construction stage, multi-source knowledge fusion technology was used to integrate the knowledge extracted from the text with the existing hydrological ontology and expert knowledge base to build a comprehensive and professional flood resilience knowledge graph. This knowledge graph not only contains static conceptual relationships, but also incorporates temporal reasoning capabilities, which can dynamically infer resilience performance under different scenarios. This embodiment realizes the ability to automatically extract flood resilience indicators from massive unstructured texts, and improves the comprehensiveness and scientificity of the indicator system. Compared with the traditional manual indicator formulation method, it not only greatly reduces the labor cost (efficiency is improved by about 80%), but also can capture the latest research progress and practical experience in a timely manner, so that the resilience assessment system can be continuously updated and optimized. In addition, based on the reasoning ability of the knowledge graph, it can also automatically discover potential new indicators and implicit relationships between indicators, providing a more in-depth and comprehensive perspective for flood resilience assessment.

[0027] This embodiment also optimizes and quantifies the flood resilience index system, improving the scientificity and operability of resilience assessment. In terms of index classification, a 4R theory implementation method based on deep graph representation learning (such as Graph Attention Networks) is introduced, which can automatically capture the complex relationship between indicators and map it to four dimensions: robustness, redundancy, resource scheduling, and rapid recovery, realizing the intelligent and precise classification of indicators (the classification accuracy is improved by about 35%). In the indicator independence test link, the time series causal discovery algorithm (such as nonlinear Granger causality test) and graph structure learning method are combined, which can not only identify the linear correlation between indicators, but also discover complex nonlinear causal relationships, effectively avoiding indicator redundancy and cross-influence, and improving the simplicity and effectiveness of the assessment system. In terms of indicator weight calculation, an adaptive weight allocation algorithm based on multi-criteria decision theory and deep reinforcement learning is developed, which can dynamically adjust the indicator weight according to the characteristics and historical data of different cities, so that the evaluation results are more in line with the actual local situation, and improve the pertinence and flexibility of the evaluation. In the construction of the indicator quantification method, the comprehensive application of fuzzy set theory and evidence theory was introduced to effectively deal with the uncertainty and ambiguity of qualitative indicators, and improve the reliability and persuasiveness of the quantitative results. A dynamic baseline value calculation method based on deep learning was introduced, which can adaptively adjust the baseline value of the indicator according to historical data and current situations, so that the resilience assessment can reflect the dynamic changes and long-term trends of the urban system. This embodiment not only realizes the scientific optimization and precise quantification of the flood resilience indicator system, but also greatly improves the automation and adaptability of the assessment process. Compared with traditional methods, it can better handle the complex relationships and dynamic changes between indicators, and the accuracy and interpretability of the assessment results are improved (the overall performance is improved by about 40%). In addition, due to the introduction of adaptive and dynamic adjustment mechanisms, the assessment system can be continuously optimized to adapt to the characteristics of different cities and changing environmental conditions, providing a strong support tool for long-term flood resilience management and decision-making.

[0028] By building a data-driven flood resilience assessment model, end-to-end automated processing from multi-source data to resilience assessment results is achieved, improving the efficiency, accuracy and scalability of the assessment. In the feature engineering stage, a spatiotemporal feature extraction method based on graph convolutional neural network (GCN) is introduced, which can simultaneously capture the spatial correlation and temporal dynamics in the urban system, effectively integrate multi-dimensional information such as geography, hydrology and socio-economy, and provide the model with richer and more representative input features. In the model selection and construction stage, an adaptive model framework based on neural architecture search (NAS) and transfer learning is developed, which can automatically build the optimal network structure according to the data characteristics and assessment requirements of different cities, and improve the generalization ability and adaptability of the model. In particular, when processing multi-scale spatiotemporal data, a multi-scale spatiotemporal attention mechanism and a dynamic graph convolutional network are used, so that the model can adaptively focus on key information at different time and space scales, improving the modeling ability of complex urban systems. During the model training process, a fast adaptation algorithm based on meta-learning and a federated learning framework were introduced, which can not only quickly learn and generalize under limited data conditions, but also realize collaborative learning of multi-city data under the premise of protecting data privacy, expanding the application scope and effect of the model. A dynamic integration strategy was adopted in the model integration stage, combining Bayesian model averaging and deep ensemble learning methods, which can automatically adjust the weights of sub-models according to different prediction tasks and time scales, improving the stability and reliability of the prediction. In addition, in order to improve the interpretability and credibility of the model, a multi-level interpretation framework based on SHAP (SHapley Additive exPlanations) value and attention mechanism was developed, which can explain the decision-making process of the model from global to local levels, providing decision makers with intuitive and reliable evaluation basis. This embodiment not only realizes the high degree of automation and intelligence of flood resilience assessment, but also improves the accuracy and reliability of the assessment. Compared with traditional statistical models, this method improves the prediction accuracy by about 50%, and performs particularly well in dealing with complex nonlinear relationships and long-term dependencies. Especially in the face of extreme events and rapidly changing urban environments, the model has shown excellent adaptability and robustness, and can adjust the assessment results in a timely manner, providing strong technical support for the dynamic management and decision-making of urban flood resilience. In addition, due to the good interpretability and visualization capabilities of the model, the assessment results are easier to be understood and accepted by decision makers and the public, which helps promote multi-party resilience management practices.

[0029] like Figure 2 As shown, according to one aspect of the present application, step S1 is further:

[0030] Step S11, collecting multi-source data of the study area, including meteorological data, hydrological data, remote sensing image data, population density data, economic data and drainage system data;

[0031] Step S12: Based on the multi-source data, outlier detection and processing are performed to obtain the multi-source data after preliminary cleaning; based on the multi-source data after preliminary cleaning, missing value processing is performed to obtain a complete data set;

[0032] Step S13: Based on the complete data set, Z-score standardization is performed to obtain a standardized data set; and the standardized data set is spatiotemporally aligned to obtain a spatiotemporally aligned data set;

[0033] Step S14: construct a spatiotemporal data cube based on the spatiotemporal aligned data set, where the dimensions of the spatiotemporal data cube include longitude, latitude, time and variables;

[0034] Step S15: Based on the spatiotemporal data cube, the data integrity index, time consistency index and space consistency index of each variable are calculated, and a data quality report is generated; based on the data quality report, the spatiotemporal data cube that meets the expected quality is obtained and stored.

[0035] According to one aspect of the present application, in step S12, outlier detection and processing are performed based on multi-source data to obtain preliminarily cleaned multi-source data, which is further:

[0036] Step S121, based on multi-source data, calculating the moving average and moving standard deviation of each time series;

[0037] Step S122: construct an abnormal threshold based on the moving average and the moving standard deviation; mark the data in the multi-source data that exceeds the abnormal threshold as a potential abnormal value;

[0038] Step S123: manually review the potential outliers to confirm that they are real outliers; use linear interpolation to replace the real outliers to obtain multi-source data after preliminary cleaning.

[0039] In one embodiment of the present application, multi-source data are collected, including meteorological data, hydrological data, remote sensing impact data, population density data, economic data and drainage system data, wherein the meteorological data include rainfall data R, temperature data T and wind speed data W, the hydrological data include river water level data H and flow data Q, the remote sensing impact data include land use type data L and impervious area ratio data I, the drainage system data include drainage network data D and flood control facilities data F, the population density data is recorded as P, and the economic data is recorded as G.

[0040] The collected data is cleaned, specifically: outlier detection and processing are performed on rainfall data R, temperature data T, wind speed data W, river water level data H, and flow data Q, to obtain the initially cleaned rainfall data R', temperature data T', wind speed data W', river water level data H', and flow data Q'. Then, missing value processing is performed on these data to obtain a complete data set.

[0041] The cleaned data is standardized, specifically: the complete data set is Z-score standardized to obtain a standardized data set. The land use type data L is uniquely encoded to obtain data L'. The impervious area ratio data I, population density data P, and economic data G are Min-Max standardized to obtain standardized impervious area ratio data I', population density data P', and economic data G'.

[0042] The standardized data are aligned in time and space. Specifically, in the time dimension, the standardized data set is unified to a 5-minute time step to obtain rainfall data R'''', temperature data T'''', wind speed data W'''', river water level data H'''', and flow data Q''''; in the spatial dimension, the data L', impervious area proportion data I', population density data P', economic data G', drainage network data D, and flood control facilities data F are resampled to a 1km×1km grid to obtain land use type data L'', impervious area proportion data I'', population density data P'', economic data G'', drainage network data D', and flood control facilities data F'.

[0043] Using all the data after spatiotemporal alignment (including rainfall data R'''', temperature data T'''', wind speed data W'''', river water level data H'''', flow data Q'''', land use type data L'', impervious area ratio data I'', population density data P'', economic data G'', drainage network data D', flood control facilities data F'), a spatiotemporal data cube C is constructed, whose dimensions include longitude x, latitude y, time t and variable v. Based on the constructed data cube C, the data integrity index (DCI), temporal consistency index (TCI) and spatial consistency index (SCI) of each variable are calculated, and a data quality report Q is generated. The constructed data cube C is stored in a distributed file system using the Apache Parquet columnar storage format. At the same time, a multidimensional index structure I is constructed, including a B+ tree index for the time dimension, an R-tree index for the space dimension and a bitmap index for the variable dimension.

[0044] In another embodiment of the present application, rainfall data R, temperature data T and wind speed data W with high temporal and spatial resolution (such as 1km×1km, 5-minute intervals) are obtained from the meteorological department; river water level data H and flow data Q are obtained from the hydrological department; land use type data L and impervious area proportion data I are obtained from remote sensing images; population density data P and GDP data G are obtained from the statistical department; drainage network data D and flood control facilities data F are obtained from the urban planning department.

[0045] Outlier detection is performed on rainfall data R, temperature data T, wind speed data W, river water level data H, and flow data Q. Specifically, the moving average (MA) and moving standard deviation (MSD) of each time series are calculated; based on MA and MSD, the anomaly threshold is calculated: Threshold = MA ± 3 * MSD; data exceeding the threshold is marked as potential outliers; potential outliers are manually reviewed to confirm the true outliers; the confirmed outliers are replaced by linear interpolation to obtain the cleaned rainfall data R', temperature data T', wind speed data W', river water level data H', and flow data Q'.

[0046] When dealing with missing values, specifically: for short-term missing values ​​(≤6 hours), linear interpolation is used to fill in; for medium-term missing values ​​(6-24 hours), similar day pattern method is used to fill in; for long-term missing values ​​(>24 hours), multiple regression model is used to fill in, and complete rainfall data R'', temperature data T'', wind speed data W'', river water level data H'', and flow data Q'' are obtained.

[0047] Data standardization is performed, specifically: Z-score standardization is performed on the complete data set: Z = (X - μ) / σ; where X is the original value, μ is the mean, and σ is the standard deviation, and the standardized rainfall data R''', temperature data T''', wind speed data W''', river water level data H''', and flow data Q''' are obtained; One-hot encoding is performed on the land use type data L to obtain the land use type data L'; Min-Max standardization is performed on the impervious area ratio data I, population density data P, and economic data G: X norm = (X - X min ) / (X max -X min ), where X max Indicates the maximum value, X min Indicates the minimum value; obtain the impervious area ratio data I', population density data P', and economic data G'.

[0048] The time series data (including rainfall data R''', temperature data T''', wind speed data W''', river water level data H''', and flow data Q''') are unified to a 5-minute time step; for data with lower resolution, cubic spline interpolation is used for temporal upsampling to obtain time-aligned rainfall data R'''', temperature data T'''', wind speed data W'''', river water level data H'''', and flow data Q''''. The spatial data (including land use type data L', impervious area ratio data I', population density data P', economic data G', drainage network data D, and flood control facility data F) are resampled to a 1km×1km grid; for vector data (drainage network data D and flood control facility data F), rasterization is used to convert them into raster data; for raster data with higher resolution, bilinear interpolation is used for downsampling; for raster data with lower resolution, bicubic interpolation is used for upsampling to obtain spatially aligned data.

[0049] Construct the spatiotemporal data cube C, specifically: define the dimensions of the data cube C: longitude x, latitude y, time t and variable v; fill all preprocessed data into the spatiotemporal data cube C according to its spatiotemporal attributes; for static data, copy and fill in the time dimension. Generate the spatiotemporal feature matrix M, specifically: perform sliding window processing on the spatiotemporal data cube C, with a window size of 5×5×12 (5km×5km×1 hour); calculate the statistical characteristics of each variable (mean, standard deviation, maximum, minimum, median, skewness, kurtosis) in each sliding window; combine the calculated features into feature vectors to form the feature matrix M.

[0050] Perform data quality assessment, specifically: calculate the data completeness index (DCI) of each variable: DCI = (total number of samples - number of missing samples) / total number of samples; calculate the temporal consistency index (TCI) of each variable: TCI = 1 - (number of time jumps / total number of samples); calculate the spatial consistency index (SCI) of each variable: SCI = 1 - (number of spatial anomaly points / total number of grids); generate a data quality report, including DCI, TCI, SCI and their spatiotemporal distribution diagrams.

[0051] The preprocessed data cube C and feature matrix M are stored in a distributed file system (such as HDFS), using the Apache Parquet column storage format to improve query efficiency; B+ tree index is used on the time dimension, R tree index is used on the space dimension, and bitmap index is used on the variable dimension to establish a multidimensional index structure.

[0052] This embodiment constructs a high-quality, multi-dimensional spatiotemporal data cube, which provides a comprehensive and reliable data foundation for subsequent flood resilience assessment. Specifically, by integrating multi-source heterogeneous data such as meteorology, hydrology, remote sensing, population, economy and drainage system, a comprehensive urban flood resilience assessment data set is constructed. This multi-dimensional data fusion can capture the complexity and dynamics of urban systems and improve the comprehensiveness and accuracy of flood resilience assessment. Using spatiotemporal alignment technology, data from different sources and different spatiotemporal resolutions are unified into a common spatiotemporal reference framework, overcoming the problem of data inconsistency in traditional methods. Data cleaning techniques are used, such as outlier detection based on moving average and standard deviation, multi-level missing value filling strategies (short-term linear interpolation, medium-term similar day pattern, long-term multivariate regression model), etc. The data quality is improved and the impact of noise and outliers on subsequent analysis is reduced. Standardization processing (such as Z-score, unique hot encoding, Min-Max standardization) further enhances the comparability and compatibility of different types of data, laying the foundation for the multi-dimensional assessment of flood resilience. The concept of spatiotemporal data cube is introduced to uniformly express multi-source data in time and space dimensions. It is not only convenient for data storage and retrieval, but also provides convenience for subsequent spatiotemporal analysis and pattern recognition. In particular, in flood resilience assessment, it can easily analyze the resilience change trend at different spatiotemporal scales. Through the sliding window technology, various statistical characteristics of each variable (such as mean, standard deviation, maximum, minimum, median, skewness, kurtosis, etc.) are calculated. It can capture the local spatiotemporal characteristics of the data, which is helpful to identify the precursors and impact range of flood events, and improve the accuracy and timeliness of resilience assessment. Indicators such as data integrity index (DCI), temporal consistency index (TCI) and spatial consistency index (SCI) are introduced to comprehensively evaluate data quality. It provides reliability guarantee for subsequent resilience assessment, making the assessment results more credible and robust. The Apache Parquet column storage format and multidimensional index structure (B+ tree, R tree and bitmap index) are adopted to improve the storage efficiency and query speed of data. This is crucial for real-time or quasi-real-time flood resilience assessment, which can quickly respond to emergencies and improve the emergency response capabilities of cities.

[0053] This embodiment builds a high-quality and efficient data processing and management system. It can not only process large-scale multi-source heterogeneous data, but also ensure the quality and consistency of the data. For flood resilience assessment, it means that more comprehensive, accurate and timely data support can be obtained, thereby improving the accuracy and timeliness of the assessment. Especially when dealing with complex urban systems, this multi-dimensional, high-quality data foundation can better capture the performance of cities in all aspects when facing flood threats, including the pressure resistance of infrastructure, the adaptability of socio-economic systems, and the resilience of ecosystems.

[0054] like Figure 3 As shown, according to one aspect of the present application, step S2 is further:

[0055] Step S21, extracting geographic identifiers of the study area based on the spatiotemporal data cube that meets the expected quality, and constructing a text corpus based on the geographic identifiers;

[0056] Step S22, performing multimodal text preprocessing on the text corpus, including word segmentation, part-of-speech tagging, named entity recognition, OCR processing, image text extraction and speech transcription, to obtain a preprocessed text dataset;

[0057] Step S23: Based on the preprocessed text data set, context-aware keyword extraction is performed to obtain a keyword set; based on the preprocessed text data set, an improved LDA algorithm and a Kalman filter are used to perform dynamic topic modeling to obtain a dynamic topic model;

[0058] Step S24: Based on the preprocessed text dataset, use the BiLSTM-CRF model, attention mechanism and hierarchical classification strategy to perform multi-granularity entity recognition; based on the results of multi-granularity entity recognition, combine the dynamic topic model to perform entity linking to obtain the final entity set;

[0059] Step S25: Based on the final entity set, semantically enhanced relationship extraction is performed using dependency syntactic analysis, remote supervised learning, and a graph neural network model to obtain a set of relationship triples;

[0060] Step S26: construct a resilience knowledge graph based on the final entity set and the relationship triple set; generate a set of candidate resilience indicators based on the resilience knowledge graph using graph algorithms and semantic analysis methods; verify and screen the set of candidate resilience indicators to obtain a preliminary resilience indicator system.

[0061] According to one aspect of the present application, in step S23, context-aware keyword extraction is performed based on the preprocessed text data set to obtain a keyword set, which is further:

[0062] Step S231, calculating the importance score of each word based on the preprocessed text data set;

[0063] Step S232: calling a preconfigured word vector model to calculate the word vector representation of each word in the preprocessed text data set; based on the word vector representation, using a sliding window technique, calculate the local context vector of each word;

[0064] Step S233: Based on the local context information, a global topic vector is calculated using topic modeling technology;

[0065] Step S234: Calculate the keyword score of each word based on the importance score, the local context vector and the global topic vector; and form a keyword set based on the keyword score.

[0066] According to one aspect of the present application, in step S23, based on the preprocessed text data set, an improved LDA algorithm and a Kalman filter are used to perform dynamic topic modeling to obtain a dynamic topic model, which is further:

[0067] Step S235: segment the preprocessed text data set by timestamp to form a predetermined number of time windows;

[0068] Step S236: In each time window, using the improved LDA algorithm to construct a topic model;

[0069] Step S237: Based on the topic model, a Kalman filter is used to track the evolution of the topic over time to obtain a dynamic topic model;

[0070] Step S238: introducing an external knowledge base to enhance the semantic understanding of the dynamic topic model, thereby obtaining an enhanced dynamic topic model.

[0071] In this embodiment, the topic model is constructed in step S236, specifically, a topic-word distribution matrix is ​​constructed, which is as follows:

[0072] Step ①: Preprocess the documents in each time window Wi, including word segmentation, stop word removal and lemma restoration, to obtain the processed document set Di'.

[0073] Construct the term frequency-inverse document frequency (TF-IDF) matrix Ti as the input of the LDA model.

[0074] Initialize the parameters of the improved LDA model, including the number of topics K (automatically determined using the topic consistency metric) and hyperparameters α and β (using an adaptive update strategy).

[0075] Step ②: Apply the improved Gibbs sampling algorithm to perform model training, including:

[0076] Read word vector information and adjust the topic assignment probability by calculating the semantic similarity of words.

[0077] Use the Hierarchical Dirichlet Process (HDP) to dynamically adjust the number of topics K to adapt to the changes in topics in different time windows. Implement an online learning mechanism to incrementally update model parameters and improve computational efficiency. Calculate the topic-word distribution matrix Φi and the document-topic distribution matrix θi.

[0078] Step ③: Apply the topic evolution detection algorithm, including: calculating the topic similarity between adjacent time windows Wi and Wi-1, identifying emerging topics, disappearing topics and persistent topics, and tracking changes in topic intensity.

[0079] Step ④: Generate the final topic-word distribution matrix Di, which contains topic evolution information.

[0080] In one embodiment of the present application, the geographic information in the data cube C is used to obtain relevant text data from academic databases, government document libraries, news websites and social media platforms to construct an original text corpus T; the original text corpus T is subjected to multimodal text preprocessing, including word segmentation, part-of-speech tagging, named entity recognition, optical character recognition (OCR) processing, image text extraction and speech transcription to obtain a preprocessed text data set T'. Based on the preprocessed text data set T', context-aware keyword extraction is performed using algorithms such as term frequency-inverse document frequency (TF-IDF), TextRank, and word vector (Word2Vec) to obtain a keyword set K. The improved LDA algorithm and Kalman filter are applied to the preprocessed text data set T' to perform dynamic topic modeling and obtain a dynamic topic model M'. Using the preprocessed text data set T', the BiLSTM-CRF model, attention mechanism and hierarchical classification strategy are used for multi-granularity entity recognition, and entity linking is performed in combination with the dynamic topic model M' to obtain the final entity set E. Based on the preprocessed text dataset T' and the final entity set E, semantically enhanced relation extraction is performed using dependency syntactic analysis, remote supervised learning, and graph neural network models to obtain a set of relation triples R. The final entity set E is used as a node, the set of relation triples R is used as an edge, and the keyword set K and the dynamic topic model M' are combined to construct a resilience knowledge graph G. Based on the resilience knowledge graph G, a set of candidate resilience indicators I is generated using graph algorithms and semantic analysis methods. The candidate resilience indicator set I is verified and screened using the data quality report Q and the predefined expert knowledge base to obtain the final resilience indicator system F.

[0081] In another embodiment of the present application, geographic identifiers (such as place names, longitude and latitude ranges) related to the study area are extracted from the data storage of step S1; using these geographic identifiers, relevant documents and reports are retrieved from academic databases (such as Web of Science, Scopus) and government document libraries through API interfaces to obtain document collection L1. Based on the geographic identifiers, web crawler technology is used to obtain text data related to flood disasters and urban resilience from news websites and social media platforms to obtain text collection L2. The document collection L1 and the text collection L2 are stored in a distributed text database (such as Elasticsearch) to form an original text corpus T.

[0082] The plain text data in the original text corpus is segmented, POS tagged and named entity recognized to obtain the preprocessed text set T1; the PDF documents in the original text corpus are processed by OCR to extract text and chart information to obtain the text image set T2; the text in the images in the text image set T2 is recognized and extracted to obtain the text set T3; the speech in the video in the original text corpus is transcribed to generate text T4; the preprocessed text set T1, text set T3 and text T4 are merged to obtain the preprocessed text dataset T'.

[0083] Based on the preprocessed text dataset, combined with TF-IDF and TextRank algorithms, the importance score of each word is calculated; the pre-trained word vector model (such as Word2Vec) is called, and the semantic similarity of words is considered to obtain the word vector representation of each word in the preprocessed text dataset; the sliding window technology is used to capture local context information and calculate the local context vector of each word; considering the global semantic information, the topic modeling technology (such as LDA) is used to calculate the global topic vector; based on the above factors, the final keyword score is calculated; the top-N words are selected according to the keyword score to form a keyword set K.

[0084] The preprocessed text dataset is divided into multiple time windows according to timestamps; in each time window, the improved LDA algorithm is used for topic modeling to obtain the topic-word distribution matrix; based on the topic-word distribution matrix, the Kalman filter is used to track the evolution of the topic over time to obtain a dynamic topic model; an external knowledge base (such as WordNet) is introduced to semantically enhance the topics in the dynamic topic model to obtain the enhanced dynamic topic model M'.

[0085] Using the BiLSTM-CRF model as the basic recognizer, the preprocessed text dataset is subjected to preliminary entity recognition to obtain entity set E1; the attention mechanism is introduced to focus on the contextual information related to resilience, and entity set E1 is optimized to obtain the optimized entity set E2; using the hierarchical classification strategy, the large-category entities in entity set E2 are first identified, and then the subcategories are subdivided to obtain the multi-level entity set E3. Using the predefined rule base, the professional terms and abbreviations in specific fields in the multi-level entity set E3 are processed to obtain entity set E4; using the enhanced dynamic topic model as the external knowledge graph, the entity linking is enhanced, and the entities in entity set E4 are linked and disambiguated to obtain the final entity set E.

[0086] The preprocessed text dataset T' is subjected to dependency syntactic analysis to obtain the syntactic relation R1 between entities; the remote supervised learning method is used to learn the relation extraction model RE using the known relations in the knowledge base as training data; the graph neural network model GNN is used to process the syntactic relation R1 to capture the long-distance dependencies between entities and obtain the relation set R2; the relation extraction model RE is applied to the relation set R2 to extract the semantic relations between entities and obtain the preliminary relation set R3; the adversarial training method is used to optimize the relation extraction model RE to improve the robustness of the model and obtain the improved relation extraction model RE'; the improved relation extraction model RE' is applied to the preliminary relation set R3 to obtain the optimized relation set R4; the active learning strategy is used to select high-information samples in the relation set R4 for manual annotation to obtain the annotated dataset L; the annotated dataset L is used to fine-tune the relation extraction model RE' to obtain the final relation extraction model RE''; the final relation extraction model RE'' is applied to the preprocessed text dataset T' and the final entity set E to extract the relations between entities and form the final relation triple set R.

[0087] Use entity set E as nodes and relationship set R as edges to construct the initial graph structure G1; use keyword set K to expand the initial graph structure G1 and add key concept nodes to obtain graph structure G2; use dynamic topic model M' to semantically enhance the nodes and edges in graph structure G2 to obtain graph structure G3; apply predefined inference rule set RS to graph structure G3 to generate implicit relationships and attributes to obtain graph structure G4; use graph embedding technologies such as TransE to calculate the vector representation of nodes and edges in graph structure G4 to obtain the final resilience knowledge graph G.

[0088] Use graph algorithms (such as PageRank algorithm) to calculate the importance score S of nodes in the resilience knowledge graph G; apply community detection algorithms (such as Louvain algorithm) to identify the core subgraph structure CS in the resilience knowledge graph G; use path analysis (such as shortest path algorithm) to analyze the semantic associations between key entities in the core subgraph structure CS to obtain the association path set P; combine the node vector representation to calculate the semantic similarity matrix SM of entities and relationships in the association path set P; combine the importance score S, core subgraph structure CS, association path set P and semantic similarity matrix SM to generate a set of candidate resilience indicators I.

[0089] Using the data quality report Q, check the quantifiability of each indicator in the candidate resilience indicator set I to obtain the initial screening indicator set I'; use the predefined expert knowledge base EK to evaluate the scientificity and representativeness of the indicators in the initial screening indicator set I' to obtain the indicator set I''; perform correlation analysis on the indicators in the indicator set I'', remove highly correlated redundant indicators, and obtain the indicator set I'''; use the sensitivity analysis method to evaluate the influence of the indicators in the indicator set I''' on the resilience assessment, and obtain the weight set W; use the fuzzy comprehensive evaluation method, combined with the indicator set I''' and the weight set W, to assign the final weight to each indicator, and obtain the final resilience indicator system F.

[0090] This embodiment uses natural language processing technology to extract knowledge and indicators related to flood resilience from a large amount of unstructured text data, and constructs a comprehensive and dynamic resilience knowledge graph. Specifically, using geographic identifiers for literature retrieval and web crawlers, the automatic collection of text data highly relevant to the study area is realized. The relevance and coverage of text data are improved, providing rich text resources for flood resilience assessment. In particular, by integrating academic literature, government reports and social media data, it can fully capture professional knowledge, policy orientation and public perception, and provide multi-angle textual basis for resilience assessment. The unified processing of texts in various formats (plain text, PDF, images, videos) expands the available information sources, especially for the extraction and utilization of non-text information such as charts and maps in flood resilience assessment, which improves the integrity and diversity of information. By combining TF-IDF, TextRank, word vector model and sliding window technology, accurate extraction of keywords related to flood resilience is achieved. Not only the word frequency and importance are considered, but also the semantic similarity and context information are incorporated, which can more accurately identify professional terms and concepts closely related to flood resilience, providing an important foundation for the subsequent construction of the indicator system. Dynamic tracking of flood resilience topics is achieved through improved LDA and Kalman filter. The ability to capture the temporal evolution of resilience concepts and concerns helps to understand the focus and changing trends of flood resilience in different periods, and provides a basis for formulating dynamic resilience assessment strategies. Accurate identification of flood resilience-related entities is achieved through the BiLSTM-CRF model, attention mechanism and hierarchical classification strategy. This multi-granularity recognition method can comprehensively capture resilience-related entities from macro to micro, such as cities, infrastructure, policy measures, etc., providing an entity foundation for building a comprehensive resilience assessment system. By combining dependency syntax analysis, remote supervised learning and graph neural networks, the complex relationships between flood resilience-related entities are extracted, which can reveal the mutual influence and dependency between different resilience elements, and help to build a more systematic and comprehensive resilience assessment model. A comprehensive flood resilience knowledge graph is constructed by integrating entities, relationships, keywords and topics. This knowledge graph not only contains static concepts and relationships, but also incorporates dynamic theme evolution, providing a strong knowledge foundation for the systematic understanding and assessment of flood resilience. Through graph algorithms, community detection, and path analysis, resilience indicators are automatically generated from knowledge graphs, which can discover potential and non-obvious resilience indicators and enrich and improve the traditional indicator system. By combining data quality assessment, expert knowledge, correlation analysis, and sensitivity analysis, the generated indicators are fully verified and screened, ensuring that the final resilience indicator system has both a theoretical basis and practical operability.

[0091] This embodiment realizes the intelligent transformation from a large amount of unstructured text data to structured resilience knowledge. Through the integration and processing of multi-source text data, a comprehensive knowledge base including academic, policy and public views is constructed to ensure the multi-angle and comprehensiveness of resilience assessment. Through dynamic topic modeling, changes in flood resilience concepts and concerns can be captured in a timely manner, so that the assessment system can be dynamically adjusted over time to maintain its timeliness and relevance. Through entity recognition, relationship extraction and knowledge graph construction, unstructured text information is converted into a structured knowledge network, providing a structured knowledge foundation for subsequent quantitative analysis and model construction. Through the indicator generation method based on the knowledge graph, new resilience indicators that may be overlooked by traditional methods can be discovered, enriching and innovating the indicator system for flood resilience assessment. The constructed knowledge graph provides a visual and explainable knowledge foundation for flood resilience assessment, which helps to understand the reasons and logic behind the assessment results. By integrating text data from multiple fields, the integration of multidisciplinary knowledge such as hydrology, urban planning, and socio-economics is achieved, providing interdisciplinary knowledge support for the comprehensive assessment of urban flood resilience. It has improved the scientificity, comprehensiveness and operability of flood resilience assessments, and provided strong knowledge support for the formulation of more accurate and effective flood prevention and control strategies.

[0092] like Figure 4 As shown, according to one aspect of the present application, step S3 is further:

[0093] Step S31: Based on the knowledge graph, a multi-head attention mechanism is used to classify the resilience indicators in the preliminary resilience indicator system based on the "4R" theory to obtain a classified indicator set;

[0094] Step S32: extracting spatiotemporal sequence data for the indicators in the classified indicator set based on the spatiotemporal data cube; performing correlation analysis and Granger causality test based on the spatiotemporal sequence data to obtain the indicator independence analysis results;

[0095] Step S33: Based on the knowledge graph, expert scoring and machine learning methods are used to perform ANP weight calculation on the classified indicator set to obtain an indicator weight set;

[0096] Step S34: Based on the indicator attributes in the classified indicator set, a quantitative function, a fuzzy evaluation system and a multi-level calculation method are constructed, and the dynamic benchmark value is calculated in combination with the historical data in the spatiotemporal data cube to form a complete indicator quantification method set;

[0097] Step S35: Based on the indicator independence analysis results, the indicator weight set and the indicator quantification method set, screening, merging and optimization are performed to obtain the final optimized resilience indicator system.

[0098] According to one aspect of the present application, step S31 is further:

[0099] Step S311: Based on the knowledge graph, each resilience indicator in the preliminary resilience indicator system is mapped to the vector space by using graph embedding technology to generate a semantic vector set;

[0100] Step S312: construct a multi-head attention mechanism, input the semantic vector set into the multi-head attention mechanism, and generate a correlation score matrix between each resilience indicator and the 4R category; the 4R category is robustness, redundancy, resource scheduling, and rapid recovery;

[0101] Step S313: Based on the correlation score matrix, use the soft voting mechanism to assign the most suitable 4R category to each resilience indicator to form a classified indicator set.

[0102] According to one aspect of the present application, step S32 is further:

[0103] Step S321, extracting spatiotemporal series data for the indicators in the classified indicator set based on the spatiotemporal data cube;

[0104] Step S322: Based on the spatiotemporal series data, calculate the correlation coefficient matrix between the indicators, including the Pearson correlation coefficient matrix, the Spearman rank correlation coefficient matrix and the mutual information matrix;

[0105] Step S323: Based on the correlation coefficient matrix, perform partial correlation analysis to obtain a partial correlation coefficient matrix between the indicators;

[0106] Step S324: Based on the spatiotemporal series data, Granger causality test is used to analyze the causal relationship between indicators and form a Granger causality matrix between the indicators;

[0107] Step S325: construct an indicator network based on the partial correlation coefficient matrix and the Granger causality matrix; use a community detection algorithm to identify highly correlated indicator clusters in the indicator network to form an indicator independence analysis result.

[0108] In this embodiment, the indicators in the indicator system F are classified based on the "4R" theory using the knowledge graph G and the multi-head attention mechanism to obtain the classified indicator set F'; based on the data cube C, the spatiotemporal series data are extracted for the indicators in the indicator set F', and the correlation analysis and Granger causality test are performed to obtain the indicator independence analysis result I; using the knowledge graph G as the basis of the network structure, combined with expert scoring and machine learning methods, the ANP weight calculation is performed on the indicator set F' to obtain the indicator weight set W'; according to the attributes of the indicators in the indicator set F', a quantitative function, a fuzzy evaluation system and a multi-level calculation method are constructed, and the dynamic benchmark value is calculated in combination with the historical data in the data cube C to form a complete indicator quantification method set QM. Based on the indicator independence analysis result I, the indicator weight set W' and the indicator quantification method set QM, the indicators are screened, merged and optimized to obtain the final optimized resilience indicator system O.

[0109] In another embodiment of the present application, all indicators are extracted from the resilience indicator system F to form an indicator set I; using the resilience knowledge graph G and graph embedding technology, each indicator is mapped to the vector space, and relevant semantic information is extracted for each indicator i∈I to generate a semantic vector set SV; a multi-head attention mechanism MHA is constructed to capture the association between the indicators and the 4R categories (robustness, redundancy, resource scheduling, and rapid recovery); the semantic vector set SV is input into the multi-head attention mechanism MHA to obtain the association score matrix AS between each indicator and the 4R category; the association score matrix AS is processed using a soft voting mechanism to assign the most suitable 4R category to each indicator to obtain the classified indicator set F'.

[0110] Using data cube C, extract the corresponding spatiotemporal series data for each indicator in the classified indicator set F' to form a data set D; calculate the Pearson correlation coefficient matrix PCM between indicators in data set D; calculate the Spearman rank correlation coefficient matrix SCM between indicators in data set D; calculate the mutual information matrix MIM between indicators in data set D; use the partial correlation analysis method to process the Pearson correlation coefficient matrix PCM, Spearman rank correlation coefficient matrix SCM and mutual information matrix MIM to obtain the modified correlation matrix CM; apply Granger causality test to data set D, analyze the causal relationship between indicators, and obtain the causal relationship matrix GCM; construct the indicator network IN based on the correlation matrix CM and the causal relationship matrix GCM. Use community detection algorithms (such as Louvain algorithm) to process the indicator network IN and identify highly correlated indicator clusters IC. Combine the correlation matrix CM, causal relationship matrix GCM and indicator cluster IC to generate the indicator independence analysis result I.

[0111] The ANP network AN is constructed by using the resilience knowledge graph G as the initialization basis of the network structure; the judgment matrix JM of the network AN is constructed by combining expert scoring and machine learning methods (such as the feature importance of random forest); the uncertainty in the judgment matrix JM is processed using fuzzy mathematics theory to obtain the fuzzy judgment matrix FJM; based on the fuzzy judgment matrix FJM, the influence relationship between the elements in the network AN is calculated to obtain the weighted super matrix WS; an adaptive weight adjustment mechanism is introduced to dynamically update the weighted super matrix WS to obtain the limit super matrix LS; the indicator weights are extracted from the limit super matrix LS to obtain the initial weight set W; the Monte Carlo simulation method is used to evaluate the stability and sensitivity of the initial weight set W to obtain the final indicator weight set W'.

[0112] Perform attribute analysis on each indicator in the indicator set F' to determine its properties (such as positive, negative, interval type) and obtain the attribute set A; construct a special quantitative function for each indicator based on the attributes in the attribute set A to form a function set Q; for the qualitative indicators in the indicator set F', construct a fuzzy evaluation system FES to convert language variables into numerical values; for the composite indicators in the indicator set F', construct a multi-level calculation method MLC to consider the relationship between sub-indicators; based on the data cube C, extract historical data sequences and calculate the dynamic benchmark value set B; construct a normalization method NM to ensure the comparability of different indicators; integrate the function set Q, the fuzzy evaluation system FES, the calculation method MLC, the dynamic benchmark value set B and the normalization method NM to form a complete indicator quantification method set QM.

[0113] Combined with the indicator independence analysis results I, the indicator weight set W' and the indicator quantification method set QM, the comprehensive importance score of each indicator is calculated to obtain the score set S; the threshold T is set, and the key indicator set KI is screened out according to the score set S; the hierarchical clustering algorithm is used to merge similar indicators in the key indicator set KI to obtain the streamlined indicator set RI; the structural equation model is applied to verify the structural validity of the streamlined indicator set RI to obtain the revised indicator set MI; the genetic algorithm is used to optimize the indicator combination in the indicator set MI with the goal of maximizing the amount of information and minimizing redundancy to obtain the final optimized resilience indicator system O.

[0114] This embodiment systematically optimizes and quantifies the resilience indicators extracted from the text, and finally forms a scientific, comprehensive and operational flood resilience assessment indicator system. Specifically, through the multi-head attention mechanism and graph embedding technology, the intelligent matching of indicators and 4R categories (robustness, redundancy, resource scheduling, and rapid recovery) is achieved. Not only the semantic information of the indicators is considered, but also the structural information in the knowledge graph is used to make the classification results more accurate and reasonable. For flood resilience assessment, this theory-based classification can ensure that the assessment system covers all key aspects of resilience and provide a more comprehensive and systematic assessment perspective. Through a variety of statistical methods (such as correlation analysis, partial correlation analysis, Granger causality test) and network analysis methods, the relationship between indicators is comprehensively analyzed. Not only can redundant indicators be identified and eliminated, but also the potential causal relationship between indicators can be revealed. For flood resilience assessment, this means that a more streamlined, efficient and information-rich indicator system can be constructed, avoiding information duplication between indicators and improving the efficiency and accuracy of the assessment. By combining expert knowledge, machine learning and fuzzy mathematics theory, a more objective and comprehensive indicator weight calculation is achieved. In particular, the introduction of adaptive weight adjustment mechanism and Monte Carlo simulation makes the weight calculation more robust and reliable. Its application in flood resilience assessment can more accurately reflect the relative importance of different indicators in overall resilience, and improve the scientificity and credibility of the assessment results. By constructing a special quantification function, fuzzy evaluation system and multi-level calculation method for each indicator, accurate quantification of various indicators is achieved. In particular, the concept of dynamic benchmark value is introduced, so that the quantitative results can reflect relative changes. Its application in flood resilience assessment can unify qualitative and quantitative indicators into a comparable scale, which is convenient for comprehensive evaluation and horizontal comparison. By comprehensively considering the importance, independence and structural validity of the indicators, and optimizing them with genetic algorithms, an indicator system with maximum information and minimum redundancy is finally formed. It ensures that the final flood resilience assessment system is both comprehensive and concise, with both theoretical basis and practical operability.

[0115] This embodiment realizes the transformation from a preliminary set of indicators to a high-quality, operational flood resilience assessment system. Through the classification based on the "4R" theory, the assessment system is closely integrated with the resilience theory. This theory-oriented approach makes the assessment results more interpretable and scientific, and can comprehensively evaluate the flood resilience level of the city from multiple dimensions (robustness, redundancy, resource scheduling, and rapid recovery). Through strict independence tests, the information overlap between indicators is eliminated, and the efficiency and accuracy of the assessment are improved. This is particularly important for flood resilience assessment because it ensures that each indicator can provide unique information and avoids the problem of over-emphasis on some aspects and neglect of others. The use of advanced weight calculation methods, combined with expert knowledge and data-driven methods, makes the determination of indicator weights more objective and scientific. The credibility and persuasiveness of flood resilience assessment results are improved, providing a more reliable basis for policy making and resource allocation. By introducing dynamic benchmark values ​​and adaptive weight adjustment mechanisms, the assessment system can adapt to the characteristics of different cities and changes over time. This dynamism is particularly important for flood resilience assessment because urban systems are constantly changing and the assessment methods also need to be able to reflect such changes. By constructing a special quantitative method, the unified quantification of various indicators is achieved, which solves the problem of difficult quantification of common qualitative indicators in flood resilience assessment and makes comprehensive assessment possible. Through system optimization, a balance is found between comprehensiveness and simplicity. This optimized indicator system can not only comprehensively reflect the flood resilience of the city, but also does not increase the complexity and cost of the assessment due to too many indicators. By establishing a set of systematic indicator optimization and quantification methods, a replicable and comparable standardized method is provided for flood resilience assessment, which is conducive to horizontal comparison and vertical tracking between different cities and provides a basis for experience exchange and learning between cities. Through clear indicator classification and quantification methods, the assessment results are easier to explain and understand, which is of great significance for conveying the assessment results to decision makers and the public, and formulating targeted improvement measures. Through the application of a series of algorithms and tools, the automation and intelligence level of the assessment process are improved. Not only the assessment efficiency is improved, but also the deviation caused by human factors is reduced. By comprehensively considering indicators and expert knowledge in multiple fields, the organic integration of multiple disciplines such as hydrology, urban planning, and social economy in flood resilience assessment is achieved, which can more comprehensively reflect the comprehensive performance of urban systems when facing flood threats. The above effects together constitute a scientific, comprehensive, operational and dynamic flood resilience assessment system. It can not only accurately assess the current flood resilience level of the city, but also guide future resilience improvement work. It provides a powerful decision-making support tool for urban managers, helps to formulate more accurate and effective flood prevention and control strategies, and ultimately improves the overall ability of the city to cope with flood disasters.

[0116] like Figure 5As shown, according to one aspect of the present application, step S4 is further:

[0117] Step S41: Based on the spatiotemporal data cube and the resilience indicator system, feature engineering is performed, including time series feature extraction, wavelet transform and autoencoder processing, to form a feature matrix; combined with the knowledge graph, semantic features are added to the feature matrix;

[0118] Step S42: Calculate the resilience index value based on the feature matrix and the index quantification method set; construct and divide the training set and the test set based on the resilience index value using stratified sampling, SMOTE algorithm and data enhancement technology;

[0119] Step S43: Use neural architecture search technology to build a resilience assessment model; based on the training set, use the Bayesian optimization method to perform hyperparameter tuning on the resilience assessment model to obtain a trained resilience assessment model;

[0120] Step S44: Based on the test set, the performance of the trained resilience assessment model is evaluated by using the Bootstrap method, including feature importance analysis and adversarial testing, to obtain a tested resilience assessment model;

[0121] Step S45: calling the tested toughness assessment model to generate a toughness assessment report.

[0122] In this embodiment, based on the data cube C and the indicator system O, feature engineering is performed, including time series feature extraction, wavelet transform and autoencoder processing, and semantic features are constructed in combination with the knowledge graph G to form the final feature matrix X; the resilience index value Y is calculated by combining the feature matrix X and the quantification method set QM. Stratified sampling, SMOTE algorithm and data enhancement technology are used to construct and divide the data set D train , D val and D test; Implement a variety of deep learning models (such as LSTM, GRU, TCN, Transformer), use neural architecture search technology to build the model structure, apply knowledge distillation, pruning and quantization technology to optimize the model, and obtain the candidate model set FM. Train the models in the candidate model set FM, use mixed precision training, gradient accumulation, adversarial training and other technologies to improve model performance, and use Bayesian optimization to tune hyperparameters to obtain the optimized model set OM; implement a variety of integration methods, such as bagging, boosting, and stacking, use genetic algorithms to optimize the combination and weights of models in the model set OM, introduce time-aware integration strategies and uncertainty estimation, and form the final resilience assessment model E; use a variety of evaluation indicators and bootstrap methods to evaluate the performance of the resilience assessment model E on the test set, perform feature importance analysis and adversarial testing, verify the model results in combination with expert knowledge, and generate the final model evaluation report MR.

[0123] In another embodiment of the present application, from the data cube C, the relevant original features are extracted according to the optimized indicator system O to form a feature set F; the sliding window method is applied to the time series data in the feature set F to calculate the statistical features (such as mean, variance, trend, periodicity) to obtain the time series feature set TF; the data in the feature set F is subjected to wavelet transform to extract multi-scale time-frequency features to obtain the frequency domain feature set FF; the knowledge graph G is used to construct semantic enhancement features for the features in the feature set F to obtain the semantic feature set SF; the feature set F is processed using an autoencoder to learn the implicit representation of the data to obtain the potential feature set LF. The feature set F, the time series feature set TF, the frequency domain feature set FF, the semantic feature set SF and the potential feature set LF are integrated to form the final feature matrix X.

[0124] Combine the feature matrix X and the indicator quantification method set QM to calculate the resilience index value and obtain the target variable Y; use the stratified sampling strategy to select representative samples from the feature matrix X and the target variable Y to form the initial data set D1; apply the SMOTE algorithm to deal with the class imbalance problem in the initial data set D1 to obtain the balanced data set D2; apply data enhancement techniques to the data set D2, such as adding noise, time shifting, etc., to increase the sample diversity and obtain the enhanced data set D3; use the K-fold cross validation method to divide the enhanced data set D3 into the training set D train , validation set D val and the test set D test .

[0125] Implement a variety of deep learning models, including LSTM, GRU, TCN, Transformer, etc., to form a model library ML; build a model search space SS, including network structure, activation function, optimizer and other hyperparameters; use neural architecture search (NAS) technology to automatically build the model structure in the model search space SS to obtain the candidate model set CM; apply knowledge distillation technology to the models in the candidate model set CM to extract the streamlined model set RM; prune and quantize the models in the streamlined model set RM to further optimize the model structure and obtain the efficient model set EM; integrate the federated learning framework to support the distributed training of models in the efficient model set EM to form the final candidate model set FM.

[0126] Implement a variety of learning rate scheduling strategies, such as cosine annealing and cyclic learning rate, to form a strategy set LRS; use mixed precision training technology to perform preliminary training on each model in the candidate model set FM to obtain the initial model set IM; implement gradient accumulation technology to support large-scale training of models in the initial model set IM to obtain the improved model set MM; apply adversarial training and data augmentation techniques to the models in the improved model set MM to improve model robustness and obtain the enhanced model set AM; use the Bayesian optimization algorithm to perform hyperparameter tuning on the models in the enhanced model set AM to obtain the optimized model set OM.

[0127] Implement a variety of integration methods, such as Bagging, Boosting, and Stacking, to form an integration method library EL; use genetic algorithms to optimize the combination and weights of models in the model set OM to obtain the preliminary integration model EM1; introduce a time-aware integration strategy to dynamically adjust the model weights in the preliminary integration model EM1 according to the prediction time scale to obtain the dynamic integration model EM2; implement the model uncertainty estimation method to provide confidence intervals for the prediction results of the dynamic integration model EM2 to obtain a model EM3 with uncertainty quantification; build an interpretability module to analyze the decision-making process of the model EM3 to form the final resilience assessment model E.

[0128] Implement multiple evaluation indicators, such as MAE, RMSE, R2, NSE, etc., to form the evaluation indicator set EI; use the Bootstrap method to estimate the resilience evaluation model E in the test set D testThe confidence interval of the performance is obtained to obtain the performance distribution PD; the feature importance analysis method is implemented to identify the key influencing factors in the resilience assessment model E and obtain the importance ranking IR; the adversarial test scenario AS is constructed to evaluate the performance of the resilience assessment model E under extreme conditions and obtain the robustness report RR; the human-machine collaborative evaluation mechanism is introduced to verify the results of the resilience assessment model E in combination with expert knowledge to form the expert feedback report ER. The performance distribution PD, importance ranking IR, robustness report RR and expert feedback report ER are integrated to generate the final model evaluation report MR.

[0129] This embodiment constructs a highly intelligent and automated data-driven flood resilience assessment model. This model can extract key features from multi-source heterogeneous data, and through advanced machine learning and deep learning technologies, it can accurately assess and predict urban flood resilience. Specifically, by combining multiple advanced technologies such as time series analysis, wavelet transform, knowledge graph and autoencoder, deep feature extraction of raw data is achieved. This multi-dimensional and multi-scale feature engineering method can capture complex patterns and implicit information related to flood resilience. In particular, when processing spatiotemporal data, it can effectively extract the time trends, periodic changes and spatial distribution characteristics related to flood resilience, providing high-quality input for subsequent model training. Through stratified sampling, SMOTE algorithm and data enhancement technology, the problems of data imbalance and sample scarcity are solved. This is particularly important for flood resilience assessment, because extreme events are usually low-frequency and easily lead to data imbalance. This embodiment ensures that the training data can fully represent various flood scenarios, improves the generalization ability of the model and the prediction ability of extreme events. By implementing multiple deep learning models and using neural architecture search (NAS) technology, automatic optimization of model structure is achieved. Not only can the model structure that best suits flood resilience assessment be found, but the model structure can also be automatically adjusted according to the characteristics of specific cities and data features, improving the adaptability and performance of the model. Advanced technologies such as mixed precision training, gradient accumulation, and adversarial training have improved the efficiency and robustness of model training. In particular, the application of Bayesian optimization has achieved automatic tuning of hyperparameters, reduced manual intervention, and improved model performance. This is particularly important for processing large-scale, high-dimensional data in flood resilience assessment, and can achieve efficient training under limited computing resources. The genetic algorithm is used to optimize model combinations and weights, and a time-aware integration strategy is introduced to achieve a more intelligent and dynamic model integration. The model weights can be dynamically adjusted according to different prediction time scales, improving the accuracy of flood resilience assessment at different time scales. At the same time, the introduction of uncertainty estimation provides confidence intervals for the assessment results and increases the reliability of decision-making. By building an interpretability module, the transparency of the model decision-making process is enhanced. This is particularly important for flood resilience assessment because it can help decision makers understand the reasons behind the assessment results and thus formulate more targeted improvement measures. The model was fully evaluated and validated through multiple evaluation indicators, Bootstrap method, feature importance analysis and adversarial testing. This multi-angle evaluation method ensures the reliability and robustness of the flood resilience assessment model, especially its performance under extreme conditions.

[0130] This embodiment implements a highly intelligent and automated flood resilience assessment model. Through deep learning and ensemble learning techniques, the model can capture complex nonlinear relationships and spatiotemporal dependencies, improving the accuracy of flood resilience assessment. This enables city managers to understand the resilience level of the city more accurately, providing a basis for precise policy implementation. The model can not only assess the current resilience level, but also predict future trends. This predictive capability is crucial for formulating forward-looking flood prevention and control strategies, and can help cities better cope with future flood risks. Through efficient feature engineering and model training techniques, the system can quickly process real-time data and realize dynamic assessment of flood resilience. This enables city managers to grasp the changes in resilience in a timely manner and respond quickly to emergencies. The model can simultaneously process data at different spatiotemporal scales, realizing multi-scale resilience assessment from micro (such as streets and communities) to macro (such as the entire city). This multi-scale assessment capability helps to formulate more refined and hierarchical flood prevention and control strategies. Through data enhancement and adversarial testing, the model enhances the ability to predict and assess extreme flood events. This is particularly important for improving the ability of cities to cope with rare disasters. The interpretability module of the model makes the assessment results no longer a "black box" output, but can provide specific analysis of influencing factors. This helps decision makers understand the underlying reasons for the resilience assessment results and formulate more targeted improvement measures. Through automated model selection and optimization, the system can adapt to the characteristics and data features of different cities. This adaptability makes the assessment method highly transferable and can be applied to different cities and regions. Efficient model training and optimization techniques improve computational efficiency, making large-scale and high-frequency resilience assessments possible. This is essential for real-time monitoring and rapid decision support. By introducing uncertainty estimates, the model can provide confidence intervals for the assessment results. This uncertainty quantification is very important for risk assessment and decision making, and can help decision makers better weigh the risks and benefits of different options. The automation and standardization of the model enables the experience and knowledge of flood resilience assessment to be systematically accumulated and passed on, which helps to form a continuously improving assessment system and continuously improve the accuracy and practicality of the assessment. The above effects together constitute an intelligent, dynamic and accurate flood resilience assessment system. It can not only accurately assess the current resilience level, but also predict future trends, providing strong decision support for urban flood prevention and control and resilience improvement. It can help urban managers better understand the vulnerability and adaptability of urban systems, formulate more scientific and effective disaster prevention and mitigation strategies, and ultimately improve the city’s overall flood resilience and reduce losses caused by flood disasters.

[0131] According to another aspect of the present application, there is also provided an efficient flood resilience assessment system based on natural language processing and data-driven models, comprising:

[0132] at least one processor; and,

[0133] a memory communicatively connected to at least one of the processors; wherein,

[0134] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the efficient flood resilience assessment method based on natural language processing and data-driven models as described in any of the above technical solutions.

[0135] The preferred embodiments of the present invention are described in detail above; however, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. An efficient flood resilience assessment method based on natural language processing and data-driven models, characterized by: The steps include: Step S1, collect multi-source data of the study area, and perform data cleaning, standardization and spatiotemporal alignment on the multi-source data to construct a spatiotemporal data cube; perform data quality assessment on the spatiotemporal data cube; obtain and store a spatiotemporal data cube that meets the expected quality, wherein the multi-source data includes meteorological, hydrological, remote sensing, population, economic and drainage system data; Step S2: extract geographic identifiers of the study area from the spatiotemporal data cube and construct a text corpus; perform multimodal text preprocessing on the text corpus, and use natural language processing technology to perform keyword extraction, dynamic topic modeling, entity recognition, and relationship extraction to obtain basic information of the graph; construct a knowledge graph of flood resilience based on the basic information of the graph, and use the knowledge graph to generate a preliminary resilience indicator system; Step S3: Based on the knowledge graph, classify the resilience indicators in the preliminary resilience indicator system, obtain the indicator set, and perform independence test and weight calculation; use the spatiotemporal data cube to extract spatiotemporal series data for the indicators, and perform correlation and causality analysis; use the indicator quantification method and use the historical data in the spatiotemporal data cube to calculate the dynamic benchmark value; obtain the final optimized resilience indicator system; the resilience indicator system includes indicator definition, classification, weight and quantification method; Step S4: Based on the resilience index system and the spatiotemporal data cube, feature engineering is performed to generate a feature matrix; semantic features are added to the feature matrix in combination with the knowledge graph; the quantification method is called to calculate the target variable; a data set containing the feature matrix and the target variable is constructed, and divided into a training set and a test set; a resilience assessment model is constructed and trained with the training set; when in use, the trained resilience assessment model is called to generate a resilience assessment report; Step S1 is further as follows: Step S11, collecting multi-source data of the study area, including meteorological data, hydrological data, remote sensing image data, population density data, economic data and drainage system data; Step S12: Based on the multi-source data, outlier detection and processing are performed to obtain the multi-source data after preliminary cleaning; based on the multi-source data after preliminary cleaning, missing value processing is performed to obtain a complete data set; Step S13: Based on the complete data set, Z-score standardization is performed to obtain a standardized data set; and the standardized data set is spatiotemporally aligned to obtain a spatiotemporally aligned data set; Step S14: construct a spatiotemporal data cube based on the spatiotemporal aligned data set, where the dimensions of the spatiotemporal data cube include longitude, latitude, time and variables; Step S15: Based on the spatiotemporal data cube, the data integrity index, time consistency index and space consistency index of each variable are calculated, and a data quality report is generated; based on the data quality report, the spatiotemporal data cube that meets the expected quality is obtained and stored; Step S2 is further as follows: Step S21, extracting geographic identifiers of the study area based on the spatiotemporal data cube that meets the expected quality, and constructing a text corpus based on the geographic identifiers; Step S22, performing multimodal text preprocessing on the text corpus, including word segmentation, part-of-speech tagging, named entity recognition, OCR processing, image text extraction and speech transcription, to obtain a preprocessed text dataset; Step S23: Based on the preprocessed text data set, context-aware keyword extraction is performed to obtain a keyword set; based on the preprocessed text data set, an improved LDA algorithm and a Kalman filter are used to perform dynamic topic modeling to obtain a dynamic topic model; Step S24: Based on the preprocessed text dataset, use the BiLSTM-CRF model, attention mechanism and hierarchical classification strategy to perform multi-granularity entity recognition; based on the results of multi-granularity entity recognition, combine the dynamic topic model to perform entity linking to obtain the final entity set; Step S25: Based on the final entity set, semantically enhanced relationship extraction is performed using dependency syntactic analysis, remote supervised learning, and a graph neural network model to obtain a set of relationship triples; Step S26: construct a resilience knowledge graph based on the final entity set and the relationship triple set; generate a candidate resilience indicator set based on the resilience knowledge graph using a graph algorithm and a semantic analysis method; verify and screen the candidate resilience indicator set to obtain a preliminary resilience indicator system; Step S3 is further as follows: Step S31: Based on the knowledge graph, a multi-head attention mechanism is used to classify the resilience indicators in the preliminary resilience indicator system based on the "4R" theory to obtain a classified indicator set; Step S32: extracting spatiotemporal sequence data for the indicators in the classified indicator set based on the spatiotemporal data cube; performing correlation analysis and Granger causality test based on the spatiotemporal sequence data to obtain the indicator independence analysis results; Step S33: Based on the knowledge graph, expert scoring and machine learning methods are used to perform ANP weight calculation on the classified indicator set to obtain an indicator weight set; Step S34: Based on the indicator attributes in the classified indicator set, a quantitative function, a fuzzy evaluation system and a multi-level calculation method are constructed, and the dynamic benchmark value is calculated in combination with the historical data in the spatiotemporal data cube to form a complete indicator quantification method set; Step S35: Screen, merge and optimize based on the indicator independence analysis results, indicator weight set and indicator quantification method set to obtain the final optimized resilience indicator system; Step S4 is further as follows: Step S41: Based on the spatiotemporal data cube and the resilience indicator system, feature engineering is performed, including time series feature extraction, wavelet transform and autoencoder processing, to form a feature matrix; combined with the knowledge graph, semantic features are added to the feature matrix; Step S42: Calculate the resilience index value based on the feature matrix and the index quantification method set; construct and divide the training set and the test set based on the resilience index value using stratified sampling, SMOTE algorithm and data enhancement technology; Step S43: Use neural architecture search technology to build a resilience assessment model; based on the training set, use the Bayesian optimization method to perform hyperparameter tuning on the resilience assessment model to obtain a trained resilience assessment model; Step S44: Based on the test set, the performance of the trained resilience assessment model is evaluated by using the Bootstrap method, including feature importance analysis and adversarial testing, to obtain a tested resilience assessment model; Step S45, calling the tested toughness assessment model to generate a toughness assessment report; In step S12, outlier detection and processing are performed based on the multi-source data to obtain the preliminarily cleaned multi-source data, which is further: Step S121, based on multi-source data, calculating the moving average and moving standard deviation of each time series; Step S122: construct an abnormal threshold based on the moving average and the moving standard deviation; mark the data in the multi-source data that exceeds the abnormal threshold as a potential abnormal value; Step S123: manually review the potential outliers to confirm that they are real outliers; use linear interpolation to replace the real outliers to obtain preliminarily cleaned multi-source data; In step S23, context-aware keyword extraction is performed based on the preprocessed text data set to obtain a keyword set, which is further: Step S231, calculating the importance score of each word based on the preprocessed text data set; Step S232: calling a preconfigured word vector model to calculate the word vector representation of each word in the preprocessed text data set; based on the word vector representation, using a sliding window technique, calculate the local context vector of each word; Step S233: Based on the local context information, a global topic vector is calculated using topic modeling technology; Step S234: Calculate the keyword score of each word based on the importance score, the local context vector and the global topic vector; and form a keyword set based on the keyword score; In step S23, based on the preprocessed text data set, an improved LDA algorithm and a Kalman filter are used to perform dynamic topic modeling to obtain a dynamic topic model, which is further: Step S235: segment the preprocessed text data set by timestamp to form a predetermined number of time windows; Step S236: In each time window, using the improved LDA algorithm to construct a topic model; Step S237: Based on the topic model, a Kalman filter is used to track the evolution of the topic over time to obtain a dynamic topic model; Step S238, introducing an external knowledge base to enhance the semantic understanding of the dynamic topic model, to obtain an enhanced dynamic topic model; Step S31 is further as follows: Step S311: Based on the knowledge graph, each resilience indicator in the preliminary resilience indicator system is mapped to the vector space by using graph embedding technology to generate a semantic vector set; Step S312: construct a multi-head attention mechanism, input the semantic vector set into the multi-head attention mechanism, and generate a correlation score matrix between each resilience indicator and the 4R category; the 4R category is robustness, redundancy, resource scheduling, and rapid recovery; Step S313: Based on the correlation score matrix, a soft voting mechanism is used to assign the most suitable 4R category to each resilience indicator to form a classified indicator set; Step S32 is further as follows: Step S321, extracting spatiotemporal series data for the indicators in the classified indicator set based on the spatiotemporal data cube; Step S322: Based on the spatiotemporal series data, calculate the correlation coefficient matrix between the indicators, including the Pearson correlation coefficient matrix, the Spearman rank correlation coefficient matrix and the mutual information matrix; Step S323: Based on the correlation coefficient matrix, perform partial correlation analysis to obtain a partial correlation coefficient matrix between the indicators; Step S324: Based on the spatiotemporal series data, Granger causality test is used to analyze the causal relationship between indicators and form a Granger causality matrix between the indicators; Step S325: construct an indicator network based on the partial correlation coefficient matrix and the Granger causality matrix; use a community detection algorithm to identify highly correlated indicator clusters in the indicator network to form an indicator independence analysis result.

2. An efficient flood resilience assessment system based on natural language processing and data-driven models, characterized by: include: at least one processor; as well as, a memory communicatively connected to at least one of the processors; wherein, The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the efficient flood resilience assessment method based on natural language processing and data-driven models as described in claim 1.

Images