A flood disaster loss assessment big data analysis method and system based on multi-source data fusion

Abstract

The application discloses a flood disaster loss evaluation big data analysis method and system based on multi-source data fusion, relates to data processing technology, and comprises the following steps: collecting multi-source heterogeneous data in real time, and constructing a unified space-time grid index based on the preprocessed multi-source heterogeneous data; extracting disaster loss sensitive features corresponding to the air network data, the space network data and the ground network data; performing cross-modal feature fusion processing on the disaster loss sensitive features, combining the unified space-time grid index, and generating fusion features carrying space-time grid codes; based on the fusion features and a flood disaster knowledge graph, combining a vulnerability curve dynamically updated by the multi-source heterogeneous data collected in real time, performing flood disaster quantitative evaluation, and generating a disaster loss evaluation result; inputting the disaster loss evaluation result into a mixed reasoning model for disaster loss reasoning, and outputting a dynamic disaster loss result. The method provided by the application can meet the requirements of flood disaster loss evaluation in the whole cycle, high precision, high credibility and high efficiency.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a big data analysis method and system for flood damage assessment based on multi-source data fusion. Background Technology

[0002] Floods are characterized by their suddenness and complex nature, making accurate and efficient damage assessment crucial for emergency command, insurance claims, and post-disaster reconstruction. Currently, the industry largely employs multi-source data integration, combining satellite remote sensing, SAR imagery, hydrological monitoring, and social media data with technologies such as knowledge graphs, deep learning, and digital twins for damage assessment. However, many shortcomings still exist.

[0003] Existing technologies mostly remain at the shallow data layer fusion level, with low preprocessing efficiency and difficulty in transforming unstructured data into structured disaster damage characteristics; disaster damage models rely on static parameters, fail to address the sample imbalance problem, and have weak generalization ability; digital twin simulation lacks accuracy and efficiency, does not support multi-scenario simulation; lacks a reliable evidence storage mechanism, the assessment process is not transparent, and it does not achieve multi-level refined output, making it unable to adapt to the needs of multiple departments, and it is also difficult to use offline on edge terminals.

[0004] Therefore, existing methods cannot meet the needs of full-cycle, high-precision, high-reliability, and high-efficiency flood damage assessment, and there is an urgent need for an improved big data analysis method for flood damage assessment based on multi-source data fusion. Summary of the Invention

[0005] This application provides a big data analysis method and system for flood disaster damage assessment based on multi-source data fusion, which is used to solve the problems of shallow fusion, low accuracy, poor timeliness, lack of traceability and difficulty in edge deployment.

[0006] This application provides an embodiment of a big data analysis method for flood damage assessment based on multi-source data fusion, comprising: Real-time acquisition of multi-source heterogeneous data, and construction of a unified spatiotemporal network index based on the preprocessed multi-source heterogeneous data, wherein the multi-source heterogeneous data includes sky network data, air network data and ground network data; Extract the disaster-sensitive features corresponding to the SkyNet data, the AirNet data, and the GroundNet data; The disaster-sensitive features are subjected to cross-modal feature fusion processing, and combined with a unified spatio-space network index, a fused feature carrying a spatio-space network code is generated. The cross-modal feature fusion processing is a unified high-dimensional mapping, feature enhancement and hierarchical deep fusion processing of the disaster-sensitive features. Based on the fusion features and flood damage knowledge graph, and combined with the vulnerability curve dynamically updated from the multi-source heterogeneous data collected in real time, flood damage assessment is performed to generate damage assessment results. The disaster damage assessment results are input into the hybrid inference model for disaster damage inference, and dynamic disaster damage results are output.

[0007] This application also proposes a big data analysis system for flood damage assessment based on multi-source data fusion, including a processor and a memory. The memory stores a computer program, and when the computer program is executed by the processor, it implements the steps of the aforementioned big data analysis method for flood damage assessment based on multi-source data fusion.

[0008] The method presented in this application has the following beneficial effects: First, it significantly improves the efficiency and quality of multi-source data fusion. Through unified spatiotemporal registration, standardized preprocessing, and cross-modal deep fusion, it achieves efficient access to multi-dimensional heterogeneous data from the sky, ground, and network, eliminating data silos and heterogeneous conflicts, thus laying a solid foundation for high-precision assessment. Second, it optimizes the accuracy and dynamism of disaster damage assessment. Relying on disaster damage knowledge graphs, dynamic vulnerability curves, and hybrid inference models, it solves the problem of imbalanced samples, reduces assessment errors, and meets the needs of refined assessment at the building level.

[0009] The above description is merely an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0010] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a schematic diagram of the basic process of the big data analysis method for flood damage assessment based on multi-source data fusion in this embodiment. Detailed Implementation

[0011] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0012] This application proposes a big data analysis method for flood damage assessment based on multi-source data fusion, such as... Figure 1 As shown, it includes the following steps: In step S101, multi-source heterogeneous data is collected, and a unified spatiotemporal network index is constructed based on the preprocessed multi-source heterogeneous data. The multi-source heterogeneous data includes sky network data, air network data, and ground network data.

[0013] It should be noted that the multi-source heterogeneous data here refers to the multi-dimensional data of the Sky-Ground Network. The Sky-Ground Network refers to a three-dimensional multi-source data acquisition and collaborative observation system adapted to flood damage assessment, integrating space-based, air-based, and ground-based networks, corresponding to six-dimensional data acquisition, as detailed below.

[0014] SkyNet (also known as space-based) data: data monitored by a network of Earth observation satellites, including but not limited to data monitored by remote sensing satellites and SAR satellites, used to acquire large-scale flood-related images and meteorological data; Airborne network (also known as airborne base) data: data monitored by an observation network composed of high-altitude platforms and drones, including but not limited to data monitored by drones and high-altitude airships, used to acquire regional high-precision imagery and LiDAR point cloud data; Ground network (also known as ground-based) data: Data from various ground-based monitoring and information network monitoring, including but not limited to ground-based hydrological IoT monitoring data (including basic geographic data, disaster-bearing body data, etc.), basic geographic databases, social sensing networks, and auxiliary assessment data, used to obtain real-time monitoring data, disaster-bearing body information, and disaster reporting data; Among them, ground hydrological IoT monitoring data includes, but is not limited to, real-time monitoring data of hydrological stations, rain gauges, water level stations, and river channels via IoT; The basic geographic database includes basic geographic data and disaster-bearing asset data, including but not limited to basic geographic data, DEM, building outlines, population, economic assets, transportation, municipal facilities data, etc. Social perception data, including but not limited to internet public opinion, short videos, and disaster reporting; In addition, it includes auxiliary assessment data, including but not limited to insurance claims and historical disaster loss cases.

[0015] In one embodiment, this application provides an adaptive access mechanism to achieve adaptive access to all types of disaster-related data, solving the problem of "data silos." This mechanism adopts metadata-driven heterogeneous data mapping technology, which can identify data types of different formats (such as images, text, numerical values, vectors, etc.) without manual intervention, automatically match the corresponding access interface, and achieve "one-time access, full availability," significantly reducing data access costs.

[0016] This application addresses the issues of heterogeneous, spatiotemporal, and scale-dependent access data by normalizing the data and constructing a unified spatiotemporal network index.

[0017] Specifically, based on preprocessed multi-source heterogeneous data, a unified spatiotemporal network index is constructed, including the following steps.

[0018] Based on geographic grids and time slices, this system automatically performs a unified coordinate system transformation on spatial data from multi-source heterogeneous data collected during preprocessing. Specifically, a unified spatiotemporal registration algorithm based on geographic grids and time slices transforms all data sources (data from different coordinate systems such as satellite remote sensing, SAR, UAVs, and LiDAR) to the same geographic coordinate system, performs time alignment on minute-level time slices, and achieves scale uniformity through resampling technology, thus solving the technical challenge of fusing data from different time periods and scales.

[0019] Based on time slicing, the system automatically aligns time-series data from multi-source heterogeneous data collected during preprocessing with time slices. This involves resampling and format standardization of data at different scales and in different formats to ensure data dimensionality uniformity. Format standardization involves converting unstructured data (disaster damage text, images) into structured disaster damage features (such as building damage level and inundation range) using a CNN convolutional neural network and a BERT semantic recognition model; and uniformly normalizing the structured data to the [0,1] interval to achieve homogeneous adaptation of multi-source data.

[0020] The target multi-source heterogeneous data is obtained by resampling, denoising, missing value repair, and outlier removal of the multi-source heterogeneous data after coordinate system transformation and time slice alignment using an adaptive algorithm. Specifically, a hybrid algorithm based on isolated forest and adaptive thresholding is employed to automatically remove outliers (such as sensor fault data and false reports), impute missing values ​​(using interpolation based on spatiotemporal correlation), and denoise the data (SAR data speckle suppression and remote sensing image dehazing), thereby improving data quality. A unified spatiotemporal network index is then constructed based on the target multi-source heterogeneous data.

[0021] In this embodiment of the application, a unified spatiotemporal network index is constructed to enable fast matching and retrieval of multi-source data within the same spatiotemporal framework. The construction method of the spatiotemporal network index is as follows.

[0022] First, determine the grid reference. Use the pre-processed unified geographic coordinate system as the spatial reference and the time reference of the time series data with a unified time granularity (e.g., 10 minutes / 1 hour). Take into account both assessment accuracy (building level) and computational efficiency, and set the grid size (e.g., 10m×10m, with the core area refined to 5m×5m).

[0023] Then, a spatiotemporal grid was created. A hierarchical grid division strategy was adopted, first dividing the area into a coarse grid, and then dividing key areas such as flood-sensitive areas and densely populated areas into fine grids; at the same time, each spatial grid was bound to a corresponding time node according to time slices, forming a two-dimensional spatiotemporal grid unit of "spatial grid + time slice". Next, grid coding is performed. A unique code (integrating spatial coordinates and timestamp information) is assigned to each spatiotemporal grid cell. The coding rules are adapted to multi-source data retrieval, ensuring that all preprocessed data (such as remote sensing images, hydrological monitoring values, and building information within a certain grid) can be quickly located through coding.

[0024] Next, data association and mapping are performed. The preprocessed six-dimensional data (satellite, UAV, hydrological, disaster-bearing bodies, etc.) are associated with the corresponding spatiotemporal grid cells according to their spatiotemporal location to establish a mapping relationship of "grid coding - multi-source data" and realize a one-to-one correspondence between data and grid cells.

[0025] Finally, the index was optimized and adapted. A dynamic index update mechanism was introduced to simultaneously adapt to the real-time access of multi-source data (such as streaming monitoring data) and the requirements for spatiotemporal grid refinement; the index query algorithm was optimized to ensure that during subsequent feature fusion and disaster assessment, multi-source data within the target spatiotemporal range can be quickly retrieved and matched through grid encoding, thereby improving overall processing efficiency.

[0026] Therefore, the spatiotemporal grid provided in this application is a grid with "hierarchical classification + spatiotemporal binding + dynamic adaptation".

[0027] In step S102, disaster-sensitive features corresponding to the sky network data, the air network data, and the ground network data are extracted. The ground network data includes surface hydrological IoT monitoring data, a basic geographic database, social sensing data, and auxiliary assessment data. The basic geographic database includes basic geographic data and disaster-bearing body asset data.

[0028] Specifically, based on the constructed unified spatiotemporal network index, disaster-sensitive features are extracted, including: extracting disaster-sensitive features corresponding to the SkyNet data, the AirNet data, the surface hydrological IoT monitoring data, the basic geographic data and disaster-bearing asset data, the social perception data, and the auxiliary assessment data.

[0029] It should be noted that the disaster-sensitive feature extraction provided in this application adopts a modal differential extraction + spatiotemporal correlation calibration + metadata-driven unstructured to structured approach. Combined with the characteristics of six-dimensional data, a differential algorithm is designed. Through spatiotemporal calibration and optimization, the accuracy, consistency and usability of the features are ensured, providing core support for subsequent cross-modal deep fusion and high-precision disaster loss assessment.

[0030] Based on the unified spatiotemporal grid index of "spatial grid + time slice" constructed in the preprocessing stage, all multi-source heterogeneous data are anchored to the corresponding spatiotemporal grid unit, ensuring that the extraction of disaster-sensitive features of various types of data is based on the same spatiotemporal benchmark, and avoiding feature misalignment caused by spatiotemporal deviation.

[0031] To address the different modal characteristics of multi-source heterogeneous data, a differential extraction algorithm is used to extract disaster-sensitive features.

[0032] Tianwang Data: Employs an improved semantic segmentation algorithm to extract core disaster-sensitive features, including but not limited to inundation range, inundation boundary, inundation area, water depth level (calculated based on the correlation between image grayscale values ​​and inundation duration), inundation dynamic change rate (time-series image comparison), and features distinguishing water bodies from non-water bodies. It focuses on capturing key disaster-related features such as breaches and inundation diffusion paths, solving the problems of traditional remote sensing extraction being susceptible to weather interference and inaccurate water depth estimation.

[0033] Airborne network (also known as airborne base) data: Combining point cloud segmentation and image interpretation methods, it extracts refined disaster-sensitive features, including but not limited to the degree of building damage (feature parameters such as wall cracks and roof collapse), terrain elevation deviation (terrain changes caused by flood erosion), degree of vegetation submersion damage, and coordinates of infrastructure (roads, bridges) breakpoints. Among them, LiDAR point cloud focuses on extracting precise quantitative features of building height and submersion depth to meet the needs of refined building-level assessment.

[0034] Surface hydrological IoT monitoring data: Employing time-series feature extraction algorithms, dynamic disaster-sensitive features are extracted, including but not limited to peak water level, average flow velocity, inundation duration, flood rise / recede rate, and flow change trend. Anomaly detection algorithms are used to screen out key disaster-causing threshold features (such as the duration of exceeding safe water levels), which are directly correlated with the degree of inundation damage to the disaster-bearing body, providing dynamic parameter support for disaster damage quantification inference.

[0035] Basic geographic data and disaster-bearing body asset data are used to extract static and semi-static disaster-sensitive features through feature screening and quantitative coding. These features include, but are not limited to, disaster-bearing body type (buildings, farmland, infrastructure), building structure type (brick-concrete, frame, etc., associated with vulnerability level), disaster-bearing body value quantification parameters, population density, disaster-bearing body spatial distribution density, and infrastructure importance level (core / general). Non-quantitative features (such as structure type) are transformed into standardized feature vectors to adapt to subsequent model calculations.

[0036] Socially perceived data employs a combination of text sentiment analysis and image target detection to extract unstructured disaster-sensitive features, including but not limited to disaster hotspots (based on public opinion and short video geographic tag aggregation), affected population size estimation features, keywords describing damage to disaster-bearing bodies (such as "collapse" and "flooded"), and features indicating the urgency of rescue needs. Through metadata-driven heterogeneous mapping, unstructured features are transformed into structured feature vectors, supplementing the disaster damage information gaps in traditional monitoring data.

[0037] The auxiliary assessment data employs statistical feature extraction and case matching algorithms to extract auxiliary disaster loss sensitive features, including but not limited to the loss rate features of similar historical floods, the correlation parameters between insurance claim amounts and damage levels, and the basic vulnerability features of disaster-bearing bodies in different regions. These features provide reference features for dynamic vulnerability curve updates and disaster loss result calibration, thereby improving assessment accuracy.

[0038] After extracting disaster-sensitive features from the multi-source heterogeneous data, the disaster-sensitive features of different modalities are first spatiotemporally aligned and calibrated based on a unified spatiotemporal space net index to remove invalid features that are spatiotemporally misaligned, ensuring spatiotemporal consistency of features. Then, a feature importance algorithm is introduced to retain core features that are highly relevant to disaster assessment and remove redundant features. At the same time, an adaptive filtering algorithm is used to eliminate noise features generated during the extraction process and improve feature purity. Finally, all optimized disaster-sensitive features are standardized into fixed-dimensional feature vectors, bound to corresponding spatiotemporal space net codes, and output to the subsequent cross-modal feature fusion module to ensure that features of different modalities and types can achieve unified high-dimensional space mapping.

[0039] The extracted disaster-sensitive features are input into the three-level nested fusion framework constructed in this application to achieve the extraction of deep complementary features from multi-source data.

[0040] In step S103, the disaster-sensitive features are subjected to cross-modal feature fusion processing, and combined with a unified spatio-space network index, a fused feature carrying a spatio-space network code is generated. The cross-modal feature fusion processing is a unified high-dimensional mapping, feature enhancement and hierarchical deep fusion processing of the disaster-sensitive features.

[0041] Specifically, a cross-modal autoencoder is employed to uniformly map the disaster-sensitive features from different input modalities to the same high-dimensional feature space, extracting unified high-dimensional features. In the high-dimensional feature space, a spatiotemporal attention mechanism algorithm is activated, combining the unified spatiotemporal net index to locate high-disaster-sensitive regions, and automatically adjusting the feature weights corresponding to these regions. The unified high-dimensional features with adjusted feature weights undergo adaptive weighted assimilation, feature concatenation enhancement, and dynamic weight decision fusion processing, and the fusion result is calibrated for confidence. The calibrated fusion result is then standardized, and combined with the corresponding spatiotemporal net encoding, a unified high-dimensional fusion feature carrying the spatiotemporal net encoding is generated and output.

[0042] It should be noted that the three-level nested fusion framework is a three-level nested fusion architecture of data layer—feature layer—decision level, which sequentially completes the adaptive weighted assimilation of multi-source features, feature splicing enhancement, and dynamic weight decision fusion, and performs confidence calibration on the fusion result.

[0043] Among them, data layer fusion: an adaptive weighted assimilation algorithm is adopted, which combines the credibility of disaster-sensitive features of each modality (such as remote sensing data resolution and sensor accuracy) to dynamically allocate fusion weights and perform preliminary fusion on the mapped high-dimensional feature vectors to solve the conflict problem between heterogeneous features (such as the difference in flooding range between remote sensing and ground monitoring).

[0044] Feature layer fusion: The features after data layer fusion are concatenated, and the disaster-causing correlation between different modal features is mined through feature interaction algorithms to extract representative enhanced fusion features. This application introduces a cross-modal autoencoder to extract features from the homogeneous data after data layer fusion, mining high-dimensional correlation feature vectors of "hydrology-inundation-disaster-bearing body" (such as the nonlinear correlation between water depth and building damage level). At the same time, a spatiotemporal spectral attention mechanism is introduced to automatically focus on high-disaster-sensitive areas such as inundation areas, breaches, transportation arteries, and key buildings, improving feature utilization and solving the technical problem that shallow fusion cannot extract deep correlations. The high-disaster-sensitive areas here include the core flood inundation area, densely populated areas, critical infrastructure distribution areas, and areas surrounding breaches. The spatiotemporal attention mechanism automatically adjusts feature weights by identifying the regional attributes corresponding to the spatiotemporal lattice encoding.

[0045] Decision-level fusion: An enhanced Bayesian network combined with a dynamic weighted voting algorithm is used to perform decision fusion on the enhanced fusion features output by the feature layer. Combined with a confidence calibration algorithm, the decision weights of each feature are dynamically adjusted (e.g., remote sensing flooding features are emphasized in the early stage of the disaster, and ground verification features are emphasized in the later stage). Feature components with large fusion deviations are removed to obtain accurate fusion results.

[0046] It should be noted that the unified high-dimensional fusion feature is a standardized feature vector containing multimodal disaster loss association information. The spatiotemporal lattice encoding bound to the feature vector is consistent with the spatiotemporal lattice encoding in the preprocessing and feature extraction stages mentioned above, to ensure the spatiotemporal consistency of the features.

[0047] By performing flood hazard quantification assessment on the fusion characteristics, assessment results are generated, which can then be adapted to the assessment needs of different flood types and regions.

[0048] In step S104, based on the fused features and the flood damage knowledge graph, and combined with the vulnerability curve dynamically updated from the multi-source heterogeneous data collected in real time, a flood damage assessment is performed to generate a damage assessment result.

[0049] This application constructs a dynamic disaster damage assessment model with strong generalization and high accuracy. Its core is to construct a hybrid reasoning model of "knowledge-driven + data-driven" based on cross-modal deep fusion features, flood disaster damage knowledge graph and dynamic vulnerability curve, so as to realize dynamic quantitative assessment of flood disaster damage.

[0050] Specifically, based on a unified spatiotemporal network index, the fused features are semantically matched with a flood damage knowledge graph to extract information on disaster-bearing bodies, vulnerability indicators, and historical damage association rules corresponding to the current assessment area and flood type. Vulnerability curves dynamically updated from real-time collected multi-source heterogeneous data are invoked, and combined with the extracted disaster-bearing body information and vulnerability indicators, the loss rate of various disaster-bearing bodies under the current flood scenario is quantitatively calculated. The loss rates of various disaster-bearing bodies are calibrated and corrected using historical damage association rules from the flood damage knowledge graph to eliminate data bias. The loss rates of various disaster-bearing bodies and corresponding disaster-bearing body asset data are summarized to calculate the total damage amount and damage level of the assessment area, generating standardized damage assessment results. These damage assessment results carry corresponding spatiotemporal network codes for subsequent damage inference in the hybrid inference model.

[0051] The method for constructing a flood damage knowledge graph includes: extracting entities, entity relationship networks, and attribute information from preprocessed multi-source heterogeneous data; constructing a flood damage knowledge graph using a graph database; wherein the flood damage knowledge graph is a knowledge graph covering flood events, disaster-bearing bodies, vulnerability indicators, and loss types; and dynamically updating entity attributes and relationships in the flood damage knowledge graph based on real-time disaster damage data, historical flood damage cases, and new historical flood damage cases not added to the historical case database.

[0052] First, focusing on the entire process of flood damage assessment, a flood damage knowledge graph covering "cause of disaster - exposure - vulnerability - loss" is constructed.

[0053] The flood damage knowledge graph consists of entities, a network of entity relationships, and attribute information. Specifically, the core entities include flood events (flood type, intensity, inundation characteristics (such as inundation range, water depth, flow velocity, etc.), disaster-bearing bodies (such as buildings, farmland, population, infrastructure, etc.), vulnerability indicators (building structure, structural type, terrain conditions, etc.), loss types (direct economic loss, indirect economic loss, asset value, etc.), and loss levels (loss rate).

[0054] The entity relationships include the disaster-causing relationship of "flood event - inundation characteristics", the correlation relationship of "disaster-bearing body - vulnerability index", the exposure relationship of "inundation characteristics - disaster-bearing body", and the quantitative relationship of "disaster-bearing body - loss".

[0055] Then, entity, relationship, and attribute information are extracted from the preprocessed multi-source heterogeneous data (auxiliary assessment data, historical disaster loss cases), and a knowledge graph is constructed using a graph database to realize semantic association and knowledge reuse between entities.

[0056] Based on historical disaster damage big data and real-time monitoring data, a dynamically updatable vulnerability curve library is constructed to replace the traditional static vulnerability curve. It is adapted to different regions, building types, and flood types (basin floods, urban waterlogging, etc.), improves the model's generalization ability, and realizes a closed loop of "data update - curve optimization - model iteration".

[0057] The dynamic update method for vulnerability curves includes: collecting historical flood damage cases, insurance claims data, and real-time multi-source heterogeneous monitoring data as update samples; extracting sample features from the update samples; and using a target algorithm to screen valid samples based on these features. The sample features include inundation characteristics (water depth, flow velocity, inundation duration), disaster-bearing body attributes (type, structure), and corresponding loss rate data. Specifically, the SMOTE algorithm is used to screen and balance the samples, removing invalid and abnormal samples and retaining valid samples that truly reflect the correlation between flood inundation and disaster-bearing body losses, thus solving the sample imbalance problem and providing high-quality data support for curve fitting.

[0058] Based on geographical region, type of disaster-bearing body, and flood type, initial vulnerability curves are fitted and core parameters are determined. Specifically, the assessment area is classified according to geographical region, type of disaster-bearing body (residential, farmland, industrial facilities, transportation facilities, etc.), and type of flood (rainstorm flood, flash flood, snowmelt flood, etc.). For each scenario, based on the selected valid samples, a corresponding initial vulnerability curve is fitted, and the core parameters of the curve (inundation depth, inundation duration, correlation coefficient between flow velocity and loss rate) are determined.

[0059] By introducing a time decay factor and combining real-time multi-source heterogeneous monitoring data (such as the inundation characteristics of the current flood and the real-time status of the disaster-bearing body), the core parameters are dynamically adjusted to weaken the weight of early historical cases and strengthen the influence of recent cases and real-time data, thereby avoiding the disconnect between static curves and actual flood scenarios and improving curve adaptability.

[0060] A vulnerability curve library is constructed and stored in a categorized manner. The fitted and calibrated vulnerability curves are classified and archived according to geographical region, disaster-bearing body type, and flood type. A real-time update mechanism is established to synchronously receive new disaster damage data and flood disaster damage knowledge graph update results, and the curve parameters are calibrated regularly to ensure the adaptability of the curves to actual disaster damage scenarios.

[0061] After each curve update, the updated vulnerability curve is verified by combining recent small flood damage verification data or historical similar case data. The deviation between the curve output loss rate and the actual loss rate is calculated. If the deviation exceeds the preset threshold, the sample screening criteria, time decay factor and parameter calibration rules are adjusted in reverse to form a closed-loop update process of "collection-screening-fitting-calibration-verification-optimization" to ensure the accuracy of the vulnerability curve.

[0062] After generating the disaster loss assessment results, a hybrid inference model is used to perform disaster loss inference, thereby determining the final dynamic disaster loss result.

[0063] In step S105, the disaster damage assessment results are input into the hybrid inference model for disaster damage inference, and dynamic disaster damage results are output.

[0064] Specifically, the input data is first preprocessed and integrated to obtain the disaster damage assessment results generated by the flood disaster quantitative assessment. Simultaneously, the unified high-dimensional fusion features carrying space-time net coding, dynamically updated vulnerability curve index parameters, and semantic association information of the flood disaster damage knowledge graph output by the cross-modal feature fusion module are retrieved. After integrating and verifying the disaster damage assessment results, the unified high-dimensional fusion features carrying space-time net coding, the vulnerability curve index parameters, and the semantic association information of the flood disaster damage knowledge graph (consistency verification, removal of invalid and abnormal data), the data is integrated according to a preset format and then input into the hybrid inference model.

[0065] Then, in response to the input of integrated and verified semantic association information detected by the hybrid reasoning model, dual-drive reasoning is started simultaneously. On the one hand, the deep learning submodule performs data-driven reasoning on the unified high-dimensional fusion features and the vulnerability index parameters, and uses the trained network parameters to perform preliminary reasoning on the disaster damage assessment results, outputting data-driven disaster damage reasoning results, including the preliminary disaster damage prediction values ​​of each spatiotemporal network unit. On the other hand, the knowledge reasoning submodule simultaneously calls the semantic association information of the knowledge graph (disaster-bearing body vulnerability rules, historical disaster damage association patterns) to perform semantic verification and correction on the data-driven reasoning results, eliminating reasoning deviations that do not conform to the association rules of the knowledge graph, and obtaining knowledge-driven disaster damage reasoning results.

[0066] Secondly, model fusion and weight calibration are performed. Through a dynamic weighted fusion mechanism, the weight ratio of the data-driven disaster loss inference results and the knowledge-driven disaster loss inference results are adjusted by combining real-time multi-source monitoring data, and calibration is performed to further correct inference bias and ensure that the inference results are synchronized with the actual flood evolution.

[0067] Finally, after the fusion calibration is completed, a standardized dynamic disaster loss result containing the disaster loss level, disaster loss amount, disaster loss development trend, and time space grid code of each time space grid unit is generated and output. The result is output in a standardized format according to a preset format and used for subsequent real-time updates, digital twin simulations, and trusted evidence storage steps.

[0068] The hybrid reasoning model is a dual-drive hybrid model that integrates deep learning data-driven reasoning and flood disaster damage knowledge graph semantic reasoning. It combines cross-modal fusion features and dynamic vulnerability curve parameters for collaborative reasoning and outputs dynamic disaster damage results.

[0069] The method for constructing the hybrid inference model includes: using unified high-dimensional fusion features output by the cross-modal feature fusion module, semantic association information from the knowledge graph, and dynamic vulnerability curve parameters as inputs, and using actual disaster damage data as labels to train the hybrid inference model. By adaptively adjusting the model parameters, the inference accuracy is optimized, addressing the evaluation bias caused by imbalanced samples and the scarcity of high-loss samples.

[0070] Among them, a hybrid inference model, SMOTE-BN-FLA, is constructed by integrating knowledge-driven and data-driven approaches. This model serves as the core inference unit of the dynamic disaster loss assessment model. The specific construction process includes: first, using the SMOTE algorithm to expand the imbalanced samples (high-loss samples are scarce) to solve the sample bias problem; second, using an enhanced Bayesian network for probabilistic inference to establish the mapping relationship between fusion features and disaster loss magnitude; and finally, using an adaptive adjustment mechanism of the loss function to dynamically optimize the model parameters and improve the assessment accuracy.

[0071] Embedded with a real-time data access interface, it synchronously receives cross-modal fusion features, real-time hydrological monitoring data, and knowledge graph update information to achieve dynamic adaptation of the model inference process—that is, based on real-time disaster changes, it dynamically calls the vulnerability curves of the corresponding region and type, adjusts the inference parameters, and outputs dynamic disaster damage results.

[0072] In this embodiment, historical flood damage case data and on-site measured data are used to verify the constructed dynamic disaster damage assessment model, ensuring that the model assessment error is controlled within 8% to meet the requirements of refined assessment.

[0073] In this embodiment, a dynamic model calibration mechanism is established. By combining real-time disaster loss feedback data, knowledge graph update results, and vulnerability curve adjustment information, the model inference parameters are calibrated regularly to improve the model's generalization ability and robustness, and adapt to different flood scenarios such as torrential rain and flash floods and urban waterlogging.

[0074] In another embodiment, the method further includes a method for making advance judgments on disaster damage using a flood evolution prediction model. Specifically, this includes: based on the preprocessed multi-source heterogeneous data, integrating meteorological forecast data (rainfall, typhoons) and historical flood evolution patterns, and combining a unified spatiotemporal network index to construct a flood evolution prediction model. The historical flood damage cases include historical flood evolution patterns, historical flood evolution data, historical disaster damage data, historical supporting data, and watershed geographical feature data. The historical flood evolution data, historical disaster damage data, historical supporting data, and dynamically updated vulnerability curves are input into the flood evolution prediction model, and the pre-disaster damage assessment results are output, providing advance support for emergency material allocation and personnel relocation.

[0075] Among them, the flood evolution prediction model can predict the flood inundation range, inundation depth, and evolution speed up to 72 hours in advance. The model can also receive real-time multi-source monitoring data feedback and dynamically calibrate prediction parameters to ensure that the prediction results are consistent with the actual flood evolution trend, thus providing support for disaster risk assessment and emergency decision-making.

[0076] When an actual flood occurs, the flood evolution prediction model and the hybrid inference model can be linked in real time. When there is a deviation between the actual flood evolution and the prediction, "expedited collection of multi-source data" (such as drone patrols of key areas and encrypted monitoring by the ground network Internet of Things) is automatically triggered. The vulnerability curve parameters and cross-modal feature weights are dynamically adjusted in sync to ensure that the assessment results match the actual disaster situation in real time.

[0077] After the flood, the system automatically compares the "pre-disaster damage assessment results, dynamic disaster damage results during the disaster, and actual verification results" to extract deviation characteristics. It then optimizes the flood evolution prediction model, cross-modal fusion algorithm, and vulnerability curve parameters in reverse, forming a closed-loop iterative mechanism of "early warning-assessment-review-optimization" to improve the accuracy of subsequent assessments and early warnings.

[0078] In another embodiment, the method further includes: updating disaster loss results in real time based on distributed streaming computing, and performing dynamic simulation and visualization through digital twins, which can solve the problem of insufficient credibility and traceability of assessment results, and realize full-process traceability and verification.

[0079] Specifically, a distributed big data processing platform is built using distributed Spark + Flink streaming computing to achieve multi-source data entering the lake in seconds, fusion and loss assessment in minutes, and disaster loss updates in minutes.

[0080] A centimeter-level digital twin base is constructed, integrating DEM terrain data, building white models, and infrastructure distribution data. Real-time assessment results are combined with the digital twin model to visualize the inundation evolution process, the distribution of vegetarian spaces, and disaster damage heat maps (regional, township, and building levels).

[0081] It supports multi-scenario simulation, including but not limited to current inundation simulation, engineering reinforcement plan, flood diversion and dispatch plan, and rapid prediction of losses from secondary disaster chains (such as landslides and waterlogging caused by floods).

[0082] In another embodiment, the method further includes: classifying and marking each spatiotemporal grid unit according to the degree of disaster damage, the importance of the geographical area, and the priority of emergency response; outputting graded disaster damage assessment results by region and level; and using blockchain or trusted timestamp technology to store and hash the entire process of data collection, preprocessing, feature fusion, vulnerability curve updating, disaster damage assessment, and inference, thereby achieving traceable and tamper-proof trusted storage throughout the entire process and ensuring the transparency of the data and calculation process.

[0083] Smart contract accounting refers to pre-setting disaster loss accounting rules, weighting coefficients, and accounting formulas. The smart contract automatically executes the disaster loss accounting process, generates standardized loss reports, reduces manual intervention, avoids accounting errors, and improves accounting efficiency.

[0084] The tiered output achieves lightweight modeling through model quantization pruning and knowledge distillation techniques, supporting edge-cloud collaborative deployment (offline assessment at the edge terminal and large-scale simulation in the cloud); it outputs tiered damage assessment results (regional, township, building, and individual building levels), adapting to the different decision-making needs of multiple departments such as emergency response, insurance, finance, and housing and construction.

[0085] In another embodiment, the method further includes: end-to-end lightweight inference and edge deployment.

[0086] Specifically, model quantization pruning and knowledge distillation enable rapid offline evaluation at edge terminals.

[0087] Output graded loss assessment results: regional level, township level, building level, and plot level (individual building level) to meet different decision-making needs.

[0088] It outputs standardized reports to support use by multiple departments, including emergency response, insurance, housing and construction, and finance.

[0089] The big data analysis method for flood damage assessment based on multi-source data fusion provided in this application effectively solves the technical problems of existing flood damage assessment methods, such as shallow fusion layer, low accuracy, poor timeliness, lack of traceability, and difficulty in edge deployment. Compared with the prior art, it has the following significant advantages: The efficiency and quality of multi-source data fusion have been significantly improved. Through unified spatiotemporal registration, standardized preprocessing, and cross-modal deep fusion, efficient access and deep fusion of six-dimensional multi-source heterogeneous data from the sky, ground, and network have been achieved. This has eliminated data silos and heterogeneous conflicts, greatly improved the utilization rate of disaster-sensitive features, and laid a solid foundation for high-precision disaster damage assessment. The dynamics and accuracy of disaster damage assessment have been greatly optimized. Relying on the flood damage knowledge graph, dynamic vulnerability curve library and hybrid reasoning model, dynamic deduction and real-time updating of disaster damage assessment have been realized, effectively solving the problems of sample imbalance and weak generalization ability, and meeting the needs of building-level refined assessment. The timeliness of the assessment has been significantly enhanced. Through distributed streaming computing and dynamic simulation of digital twins, multi-source data can be entered into the lake in seconds and disaster damage results can be updated in minutes. It is adapted to the decision-making needs of the "golden 6 hours" in disaster emergency response and can quickly support emergency command, rescue and dispatch work. The credibility and traceability of the assessment results have been significantly improved. Through the trusted evidence storage mechanism of blockchain + smart contracts, the data source, integration process and assessment results are made immutable and verifiable throughout the entire chain, which enhances the credibility of disaster loss assessment and can effectively support the collaborative application of multiple departments such as insurance claims and financial compensation. The practicality and adaptability have been greatly improved. Lightweight deployment at the edge is achieved through model quantization pruning and knowledge distillation. It can achieve rapid offline assessment under extreme conditions such as no network or disaster sites. At the same time, it is adapted to various flood types such as river basin floods, urban waterlogging, and flash floods, achieving full coverage in urban and rural areas and adaptability to multiple scenarios. The overall technical process forms a closed loop, realizing full-cycle disaster loss assessment from "data acquisition - preprocessing - feature fusion - dynamic evaluation - real-time simulation - reliable output". This significantly reduces the cost of manual intervention, improves assessment efficiency and automation level, and provides a set of efficient, accurate, reliable and universal solutions for flood disaster loss assessment.

[0090] This application also proposes a big data analysis system for flood damage assessment based on multi-source data fusion, including a processor and a memory. The memory stores a computer program, which, when executed by the processor, implements the steps of the aforementioned big data analysis method for flood damage assessment based on multi-source data fusion.

[0091] Furthermore, although exemplary embodiments have been described herein, their scope includes any and all embodiments based on this disclosure that have equivalent elements, modifications, omissions, combinations (e.g., schemes involving intersections of various embodiments), adaptations, or changes. They are not limited to the examples described in this specification or during the implementation of this application, and such examples are to be construed as non-exclusive.

[0092] The above description is intended to be illustrative and not restrictive. For example, the above examples (or one or more of them) can be used in combination with each other. Other embodiments can be used by those skilled in the art when reading the above description.

[0093] The above embodiments are merely exemplary embodiments of this disclosure. Those skilled in the art can make various modifications or equivalent substitutions to this invention within the scope of the disclosure, and such modifications or equivalent substitutions should also be considered to fall within the protection scope of this invention.

Claims

1. A big data analysis method for flood damage assessment based on multi-source data fusion, characterized in that, include: Real-time acquisition of multi-source heterogeneous data, and construction of a unified spatiotemporal network index based on the preprocessed multi-source heterogeneous data, wherein the multi-source heterogeneous data includes sky network data, air network data and ground network data; Extract the disaster-sensitive features corresponding to the SkyNet data, the AirNet data, and the GroundNet data; The disaster-sensitive features are subjected to cross-modal feature fusion processing, and combined with a unified spatio-space network index, a fused feature carrying a spatio-space network code is generated. The cross-modal feature fusion processing is a unified high-dimensional mapping, feature enhancement and hierarchical deep fusion processing of the disaster-sensitive features. Based on the fusion features and flood damage knowledge graph, and combined with the vulnerability curve dynamically updated from the multi-source heterogeneous data collected in real time, flood damage assessment is performed to generate damage assessment results. The disaster damage assessment results are input into the hybrid inference model for disaster damage inference, and dynamic disaster damage results are output.

2. The big data analysis method for flood damage assessment based on multi-source data fusion as described in claim 1, characterized in that, The construction of a unified spatiotemporal network index based on preprocessed multi-source heterogeneous data includes: Based on geographic grids and time slices, the spatial data in multi-source heterogeneous data collected during the preprocessing process is automatically transformed to a unified coordinate system. Based on time slicing, the time-series data in the multi-source heterogeneous data collected during the preprocessing process are automatically aligned with time slices. The multi-source heterogeneous data that has undergone coordinate system transformation and time slice alignment is resampled, denoised, missing value repaired and outlier removed to obtain the target multi-source heterogeneous data. Based on the target multi-source heterogeneous data, a unified spatiotemporal network index is constructed.

3. The big data analysis method for flood damage assessment based on multi-source data fusion as described in claim 1, characterized in that, The process of performing cross-modal feature fusion processing on the disaster-sensitive features, combined with a unified spatiotemporal network index, generates fused features carrying spatiotemporal network encoding, including: A cross-modal autoencoder is used to uniformly map the disaster-sensitive features of different input modalities to the same high-dimensional feature space and extract unified high-dimensional features; In the high-dimensional feature space, the spatiotemporal attention mechanism algorithm is activated, and the high-disaster-damage sensitive area is located by combining the unified spatiotemporal network index, and the feature weights corresponding to the high-disaster-damage sensitive area are automatically adjusted. The unified high-dimensional features after adjusting feature weights are subjected to adaptive weighted assimilation, feature concatenation enhancement, and dynamic weight decision fusion processing, and the fusion result is calibrated for confidence. The calibrated fusion result is standardized and combined with the corresponding spatiotemporal net code to generate a unified high-dimensional fusion feature carrying the spatiotemporal net code, which is then output.

4. The flood damage assessment big data analysis method based on multi-source data fusion as described in claim 1, characterized in that, The flood damage assessment is performed based on the fused features and flood damage knowledge graph, combined with the vulnerability curve dynamically updated from the real-time collected multi-source heterogeneous data, to generate a flood damage assessment result, including: Based on the unified spatiotemporal network index, the fused features are semantically associated and matched with the flood damage knowledge graph to extract information on the disaster-bearing bodies, vulnerability indicators, and historical damage association rules corresponding to the current assessment area and flood type. The vulnerability curve, dynamically updated from real-time collected multi-source heterogeneous data, is invoked, and combined with the disaster-bearing body information and the vulnerability index, to quantitatively calculate the loss rate of various disaster-bearing bodies under the current flood scenario. By combining historical disaster loss association rules in the flood disaster loss knowledge graph, the loss rates of various disaster-bearing entities are calibrated and corrected; By summarizing the loss rates of various disaster-bearing entities and their corresponding asset data, the total disaster loss amount and disaster loss level of the assessment area are calculated, and a standardized disaster loss assessment result is generated, wherein the disaster loss assessment result carries a corresponding time-space network code.

5. The big data analysis method for flood damage assessment based on multi-source data fusion as described in claim 4, characterized in that, The dynamic update method for the vulnerability curve includes: Historical flood damage cases, insurance claims data, and real-time multi-source heterogeneous monitoring data are collected as update samples. Sample features are extracted from the update samples, and a target algorithm is used to screen effective samples based on the sample features. The sample features include inundation characteristics, disaster-bearing body attributes, and corresponding loss rate data. Based on geographical region, type of disaster-bearing body, and type of flood, initial vulnerability curves are fitted and core parameters are determined. A time decay factor is introduced, and the core parameters are dynamically adjusted in conjunction with real-time multi-source heterogeneous monitoring data. Construct a vulnerability curve library with categorized storage, establish a real-time update mechanism, synchronously receive new disaster damage data and flood damage knowledge graph update results, and periodically calibrate curve parameters.

6. The big data analysis method for flood damage assessment based on multi-source data fusion as described in claim 1, characterized in that, The process of inputting the disaster damage assessment results into the hybrid inference model for disaster damage inference and outputting dynamic disaster damage results includes: After integrating and verifying the disaster damage assessment results, unified high-dimensional fusion features carrying time-space network codes, vulnerability curve index parameters, and semantic association information of the flood disaster damage knowledge graph, the results are input into the hybrid reasoning model. In response to the input of integrated and verified semantic association information detected by the hybrid reasoning model, dual-drive reasoning is started synchronously. The deep learning submodule performs data-driven reasoning on the high-dimensional fusion features and the vulnerability index parameters to obtain data-driven disaster damage reasoning results. At the same time, the knowledge reasoning submodule calls the disaster-bearing body vulnerability rules and historical disaster damage association patterns in the knowledge graph to perform semantic verification on the data-driven disaster damage reasoning results to obtain knowledge-driven disaster damage reasoning results. The weight ratio of the data-driven disaster loss reasoning result and the knowledge-driven disaster loss reasoning result is adjusted and calibrated by using a dynamic weighted fusion mechanism that combines real-time multi-source monitoring data. After the fusion calibration is completed, disaster damage results are generated and output. These results are standardized dynamic disaster damage results that include the disaster damage level, disaster damage amount, disaster damage development trend, and time-space network code for each time-space network unit.

7. The big data analysis method for flood damage assessment based on multi-source data fusion as described in claim 1, characterized in that, The method further includes: Based on the preprocessed multi-source heterogeneous data, meteorological forecast data and historical flood evolution patterns are integrated, and a flood evolution prediction model is constructed by combining a unified spatiotemporal network index. The historical flood damage cases include historical flood evolution patterns, historical flood evolution data, historical damage data, historical supporting data, and watershed geographical feature data. The historical flood evolution data, historical disaster damage data, historical supporting data, and dynamically updated vulnerability curves are input into the flood evolution prediction model, and the disaster damage assessment results are output.

8. The big data analysis method for flood damage assessment based on multi-source data fusion as described in claim 1, characterized in that, The method also includes: updating disaster damage results in real time and performing dynamic simulation and visualization through digital twins.

9. The big data analysis method for flood damage assessment based on multi-source data fusion as described in claim 1, characterized in that, The method further includes: classifying and marking each spatiotemporal grid unit according to the degree of disaster damage, the importance of the geographical area and the priority of emergency response, outputting the graded disaster damage assessment results, and reliably storing the data of the entire disaster damage assessment process.

10. A big data analysis system for flood damage assessment based on multi-source data fusion, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, and the computer program, when executed by the processor, implements the steps of the big data analysis method for flood damage assessment based on multi-source data fusion as described in any one of claims 1 to 9.

Images