A multispectral remote sensing disaster monitoring and early warning system using deep learning

Abstract

This invention discloses a multispectral remote sensing disaster monitoring and early warning system utilizing deep learning. The system includes: a spectral data acquisition and preprocessing module, used to continuously record image data from different time periods based on surface reflectance information to obtain a raw dataset, and preprocess the raw dataset to obtain a clear image dataset; a pattern recognition module, used to classify patterns in the clear image dataset to obtain potential disaster-related feature combinations; a disaster risk assessment module, used to perform time-series analysis on potential disaster-related feature combinations to obtain probability assessment values ​​associated with disaster risk; a disaster risk early warning module, which constructs a quantitative index system for disaster risk association based on the probability assessment values, and determines early warning reference standards based on the constructed quantitative index system for disaster risk association; and a closed-loop monitoring module, which continuously monitors the real-time acquired multispectral data based on the early warning reference standards to generate early warning information output including risk levels.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing monitoring and disaster early warning technology, and in particular relates to a multispectral remote sensing disaster monitoring and early warning system utilizing deep learning. Background Technology

[0002] Multispectral remote sensing technology plays a crucial role in disaster monitoring and early warning. By capturing spectral information of different bands on the Earth's surface, it provides important data support for the prediction and management of natural disasters.

[0003] However, many current disaster monitoring methods have significant shortcomings in data analysis and early warning mechanisms. Traditional methods often rely too heavily on data from a single point in time and lack the ability to continuously track dynamic changes. More importantly, existing technologies perform poorly when dealing with data interference in complex environments. For example, factors such as cloud cover and changes in land cover often lead to misjudgments or missed reports, making it difficult to meet the needs of high-precision early warning.

[0004] Against this backdrop, the core challenges facing this field are becoming increasingly apparent. First, the dynamic nature of multispectral data requires systems to accurately capture subtle differences in spectral reflectance across different time periods. However, these differences are often influenced by various external factors, increasing the difficulty of identification. A deeper issue lies in how to correlate these differences with the potential risks of disasters, forming a scientifically sound early warning standard. If the problems of difference identification and risk correlation cannot be effectively solved, the timeliness and accuracy of the early warning system will be significantly compromised, making it difficult to provide effective warnings before disasters occur.

[0005] Therefore, how to accurately identify abnormal change patterns in multispectral data through deep learning technology and establish a scientific and reasonable early warning threshold system based on spectral differences has become a key issue in building an efficient early warning system for disasters. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention proposes a multispectral remote sensing disaster monitoring and early warning system utilizing deep learning, thereby resolving the issues present in the existing technologies.

[0007] To achieve the above objectives, the present invention provides a multispectral remote sensing disaster monitoring and early warning system utilizing deep learning, comprising:

[0008] The spectral data acquisition and preprocessing module is used to acquire surface reflection information from multiple bands, continuously record image data of different time periods based on the surface reflection information to obtain a raw dataset, and preprocess the raw dataset to obtain a clear image dataset.

[0009] The pattern recognition module is used to perform pattern classification on the clear image dataset to obtain potential disaster-related feature combinations;

[0010] The disaster risk assessment module is used to perform time series analysis on the combination of potential disaster-related features to obtain a probability assessment value associated with disaster risk;

[0011] The disaster risk early warning module is used to construct a quantitative indicator system for disaster risk association based on the probability assessment value, and to determine early warning reference standards based on the constructed quantitative indicator system for disaster risk association.

[0012] The closed-loop monitoring module is used to continuously monitor the real-time acquired multispectral data based on the aforementioned early warning reference standard and generate early warning information output including risk levels.

[0013] Optionally, the multispectral data acquisition and preprocessing module includes a multispectral data acquisition unit and a data preprocessing unit;

[0014] The multispectral data acquisition unit is used to acquire surface reflection information from multiple bands and continuously record image data at different time periods to form an original dataset containing dynamic change characteristics.

[0015] The data preprocessing unit is used to perform denoising and correction operations on each frame of the original dataset using image preprocessing techniques to generate an optimized, clear image dataset.

[0016] Optionally, the data preprocessing unit includes: a first preprocessing subunit, a second preprocessing subunit, a third preprocessing subunit, a fourth preprocessing subunit, and a fifth preprocessing subunit;

[0017] The first preprocessing subunit is used to perform preliminary denoising on each frame of image to eliminate external interference and obtain the first image dataset after denoising.

[0018] The second preprocessing subunit is used to perform geometric and radiometric correction on the denoised first image dataset to generate a corrected second image dataset;

[0019] The third preprocessing subunit uses the comparison threshold range method to perform secondary denoising on the corrected second image dataset to obtain the optimized third image dataset.

[0020] The fourth preprocessing subunit is used to group the optimized third image dataset according to time order to obtain the grouped fourth image dataset, and extract the image change patterns of the grouped fourth image dataset in each time period to obtain a set of change features.

[0021] The fifth preprocessing subunit is used to classify outlier segments of the change feature set using the random forest algorithm to obtain an optimized clear image dataset.

[0022] Optionally, the pattern recognition module includes a spectral feature extraction unit and a pattern classification unit;

[0023] The spectral feature extraction unit is used to extract the spectral reflectance characteristics of the clear image dataset, construct a multi-dimensional feature matrix containing band differences, and analyze the variation law of spectral reflectance differences in different time periods based on the multi-dimensional feature matrix to obtain feature distribution results that can reflect surface changes.

[0024] The pattern classification unit is used to classify patterns of spectral reflectance differences based on feature distribution results using a deep learning model, and to identify potential disaster-related feature combinations.

[0025] Optionally, the pattern classification unit includes: a data standardization and preliminary extraction subunit, an abnormal fluctuation separation and labeling subunit, a feature fusion and clustering subunit, and a disaster feature identification and output subunit;

[0026] The data standardization and preliminary extraction subunit is used to standardize the spectral reflectance difference data in the feature distribution results to obtain a standardized feature base; and based on the standardized feature base, a convolutional neural network model is used to perform preliminary extraction of the spectral reflectance difference patterns to determine the distribution characteristics of the difference patterns.

[0027] The abnormal fluctuation separation and labeling subunit is used to separate the abnormal fluctuation part in the differential pattern distribution characteristics using a feature filtering method to obtain a set of separated abnormal fluctuations; and to judge the fluctuation amplitude in the set of abnormal fluctuations according to a preset threshold range to obtain a labeled fluctuation dataset.

[0028] The feature fusion and clustering subunit is used to fuse the labeled fluctuation dataset with the environmental background data, determine the fused background fluctuation characteristics, and, based on the fused background fluctuation characteristics, use a clustering tool to group the background fluctuation characteristics to determine the feature combination distribution related to disaster risk.

[0029] The disaster feature identification and output subunit is used to perform statistical analysis on the distribution of the feature combinations related to disaster risk and output potential disaster-related feature combinations.

[0030] Optionally, the disaster risk assessment module includes: a feature analysis and time series analysis unit, and a risk probability assessment unit;

[0031] The feature analysis and time series analysis unit is used to perform weighted processing on the combination of potential disaster-related features to obtain a weighted feature distribution;

[0032] The risk probability assessment unit is used to calculate the probability assessment value associated with disaster risk based on the weighted feature distribution, and to determine the potential disaster risk level of the current surface changes.

[0033] Optionally, the expression for calculating the disaster risk probability by the risk probability assessment unit is:

[0034]

[0035] In the formula, W(ΔX) is the dynamic weight matrix, CNN(·) is the trained convolutional neural network, P is the disaster risk probability, and X is the original multispectral time series data.

[0036] Optionally, the disaster risk early warning module includes: a quantitative indicator construction unit, a dynamic early warning threshold setting unit, and an early warning information generation and output unit;

[0037] The quantitative indicator construction unit is used to construct a quantitative indicator system for disaster risk association based on the probability assessment value.

[0038] The dynamic early warning threshold setting unit is used to set a dynamically adjustable early warning threshold system based on different types of disaster characteristics, application scenarios, and risk levels.

[0039] The early warning information generation and output unit is used to analyze real-time monitoring data based on the early warning threshold system. When the analysis result matches the abnormal change pattern, an early warning signal is triggered, and an early warning information output containing the risk level is generated.

[0040] Compared with the prior art, the present invention has the following advantages and technical effects:

[0041] This invention relates to a multispectral remote sensing disaster monitoring and early warning system utilizing deep learning. Through a multispectral data acquisition and preprocessing module, it can acquire and continuously record surface reflectance information from multiple bands, while simultaneously denoising and correcting the raw data to generate a clear image dataset, effectively eliminating the influence of external interference factors. A pattern recognition module uses a deep learning model to classify patterns of spectral reflectance differences, accurately identifying potential disaster-related feature combinations, performing particularly well in complex environments. A disaster risk assessment module performs time-series analysis of abnormal change patterns, combining historical data to obtain probability assessment values ​​associated with disaster risk, providing a scientific basis for early warning. A disaster risk early warning module constructs a quantitative indicator system based on the probability assessment values ​​and dynamically adjusts the early warning threshold to ensure the accuracy and adaptability of the warnings. A closed-loop monitoring module automatically adjusts the data acquisition frequency based on early warning information, focusing on high-risk areas to form a closed-loop mechanism and promptly capture risk changes. Overall, this invention significantly improves the accuracy and timeliness of disaster monitoring and early warning, providing strong technical support for disaster prevention and control. Attached Figure Description

[0042] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0043] Figure 1 This is a structural diagram of a multispectral remote sensing disaster monitoring and early warning system utilizing deep learning, according to an embodiment of the present invention. Detailed Implementation

[0044] It should be noted that, unless otherwise specified, the embodiments and features described in this application are combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0045] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0046] like Figure 1 As shown in the figure, this embodiment provides a multispectral remote sensing disaster monitoring and early warning system utilizing deep learning, including: a multispectral data acquisition and preprocessing module, a pattern recognition module, a disaster risk assessment module, a disaster risk early warning module, and a closed-loop monitoring module.

[0047] The multispectral data acquisition and preprocessing module is used to acquire raw datasets through a multispectral remote sensing image acquisition system, and to perform denoising and correction operations on the raw datasets to eliminate the influence of external interference factors such as cloud cover, generating optimized and clear image datasets.

[0048] The multispectral data acquisition and preprocessing module includes a multispectral data acquisition unit and a data preprocessing unit. The multispectral data acquisition unit acquires surface reflectance information from multiple bands using a multispectral remote sensing image acquisition system, continuously recording image data for different time periods to form a raw dataset containing dynamic changes. The data preprocessing unit, based on the time-series image set, uses image preprocessing techniques to denoise and correct each frame of the image, eliminating the influence of external interference factors such as cloud cover, generating an optimized, clear image dataset, and determining clean data suitable for subsequent analysis.

[0049] Further, the data preprocessing unit includes: a first preprocessing subunit, a second preprocessing subunit, a third preprocessing subunit, a fourth preprocessing subunit, and a fifth preprocessing subunit; wherein, the first preprocessing subunit is used to perform preliminary denoising processing on each frame of image to eliminate external interference and obtain a denoised first image dataset; the second preprocessing subunit is used to perform geometric and radiometric correction on the denoised first image dataset to generate a corrected second image dataset; the third preprocessing subunit uses a contrast threshold range method to perform secondary denoising processing on the corrected second image dataset to obtain an optimized third image dataset; the fourth preprocessing subunit is used to group the optimized third image dataset according to time order to obtain a grouped fourth image dataset, and extract the image change patterns of the grouped fourth image dataset in each time period to obtain a change feature set; the fifth preprocessing subunit uses a random forest algorithm to classify the change feature set into outlier segments to obtain an optimized clear image dataset.

[0050] In a specific implementation of this embodiment, the original dataset is processed to obtain data for each frame of image. Image preprocessing techniques are used to perform preliminary denoising on each frame of image to eliminate the influence of external interference, resulting in a denoised first image dataset. Based on the denoised first image dataset, geometric and radiometric correction methods are applied to the images to address interference factors such as cloud cover, generating a corrected second image dataset. For the corrected second image dataset, feature information of clear images is obtained. By comparing with a preset threshold range, if the clarity of a certain frame of image is lower than the threshold, a second denoising process is performed to obtain an optimized third image dataset. Based on the optimized third image dataset, a segmented extraction method is used to group the image data according to time sequence, determining a grouped fourth image dataset. Using the grouped fourth image dataset, time series analysis tools are used to extract the image change patterns within each time period based on the dynamic characteristics of each group of images, obtaining a set of change features. Based on the set of change features and the attributes of clean data, if the change features of a certain time period deviate from a preset range, it is marked as an abnormal segment, generating a marked fifth image dataset. Based on the labeled fifth image dataset, the random forest algorithm is used to classify outlier segments, determine their distribution patterns, and finalize the analysis image dataset.

[0051] Furthermore, when processing time-series image datasets, for the acquisition of data for each frame, surface information is captured by a multispectral sensor within a specific time period. Assuming data is collected once a day, a time-series dataset is formed over 30 consecutive days. In the initial denoising process, a mean filtering method can be used. For noise points in the image, such as pixel anomalies caused by equipment shake, a 3x3 filtering window is set, and the value of the center pixel is replaced with the average value of the surrounding pixels, thereby smoothing the image and reducing the impact of external interference.

[0052] When correcting for interference factors such as cloud cover, geometric correction is used to adjust the spatial position of the image to ensure that images from different time periods are aligned to the same geographic coordinate system. At the same time, radiometric correction can adjust the brightness value of the image to a consistent range by referencing standard reflectance data. For example, the reflectance value of a certain band can be corrected from the original 0.2-0.8 range to the 0.1-0.7 range to eliminate the influence of atmospheric scattering.

[0053] In the sharpness threshold comparison and secondary denoising process, a sharpness threshold of 0.75 is set. If the sharpness of a certain frame is only 0.6, a secondary processing method is used, selecting the median of pixels within a 5x5 window to replace outliers, thereby improving the visibility of image details.

[0054] When extracting time-series characteristics, the 30-day image data was divided into 6 groups of 5 days each to facilitate subsequent analysis of short-term trends within each group. When extracting dynamic features, time-series analysis tools were used to analyze the reflectance changes in each group of images. For example, if the reflectance value of a certain band in a group of images gradually increased from 0.3 to 0.5 from day 1 to day 5, it indicates that the surface may have experienced vegetation growth or increased moisture. For anomaly labeling, if the reflectance value change exceeds a preset range (e.g., the normal fluctuation range is 0.1, but a certain segment reaches 0.3), it is marked as an anomaly segment. In the classification and processing of anomaly segments, a random forest algorithm was used to perform multi-dimensional analysis of the characteristics of anomaly segments, such as combining reflectance value, time period, and band information to determine whether the anomaly was caused by seasonal changes or a sudden event.

[0055] The pattern recognition module classifies spectral reflectance differences into patterns to identify anomalous change patterns. The module includes a spectral feature extraction unit and a pattern classification unit. The spectral feature extraction unit extracts spectral reflectance difference features for each band in a clear image dataset, constructs a multi-dimensional feature matrix, analyzes the changing patterns of spectral reflectance differences over different time periods, and obtains feature distribution results that reflect surface changes. The pattern classification unit, based on the feature distribution results, uses a deep learning model to classify spectral reflectance differences into patterns, identifying anomalous change patterns, especially anomalous fluctuations under complex environmental influences, and determining potential disaster-related feature combinations.

[0056] As a specific implementation method of this embodiment, for a clear image dataset, spectral analysis tools are used to extract the spectral reflectance characteristics of each band, constructing a multi-dimensional feature matrix containing band differences to obtain a preliminary spectral reflectance feature set. Based on the preliminary spectral reflectance feature set, a time-series segmentation method is used to group the feature matrix according to the differences within a time period, determining the spectral reflectance change trend within different time periods. For the spectral reflectance change trend within different time periods, relevant feature information of land surface changes is obtained. By comparing with a preset threshold range, if the change trend of a certain time period exceeds the threshold range, it is marked as an anomalous change segment, obtaining a marked feature distribution set. Based on the marked feature distribution set, a random forest algorithm is used to classify the anomalous change segments according to their distribution characteristics, determining the distribution pattern of the anomalous change segments, and obtaining the classified feature distribution results. Based on the classified feature distribution results, the dynamic characteristics of land surface changes are obtained, and the change patterns within each time period are extracted using time-series analysis tools to determine the feature distribution pattern of land surface changes. Based on the feature distribution pattern of land surface changes, and considering the overall characteristics of the image dataset, a data integration method is used to summarize the feature distribution, obtaining the final land surface change feature set. For the final set of surface change characteristics, the distribution characteristics related to band differences are obtained. By comparing and analyzing the distribution sets in different time periods, the long-term trend characteristics of surface change are determined.

[0057] As a specific implementation of this embodiment, the pattern classification unit includes: a data standardization and preliminary extraction subunit, an abnormal fluctuation separation and labeling subunit, a feature fusion and clustering subunit, and a disaster feature identification and output subunit; wherein, the data standardization and preliminary extraction subunit is used to standardize the spectral reflectance difference data in the feature distribution results to obtain a standardized feature basis; and based on the standardized feature basis, a convolutional neural network model is used to perform preliminary extraction of spectral reflectance difference patterns to determine the distribution characteristics of the difference patterns; the abnormal fluctuation separation and labeling subunit is used to use a feature screening method to identify abnormal fluctuations in the distribution characteristics of the difference patterns. The system is divided into several parts to obtain a set of abnormal fluctuations. The amplitude of fluctuations in the set is then judged according to a preset threshold range to obtain a labeled fluctuation dataset. The feature fusion and clustering subunit is used to fuse the labeled fluctuation dataset with environmental background data to determine the fused background fluctuation characteristics. Based on the fused background fluctuation characteristics, a clustering tool is used to group the background fluctuation characteristics to determine the distribution of feature combinations related to disaster risk. The disaster feature identification and output subunit is used to perform statistical analysis on the distribution of feature combinations related to disaster risk and output potential disaster-related feature combinations.

[0058] Furthermore, when processing feature distribution data, standardization of spectral reflectance information involves using data preprocessing tools to adjust reflectance values ​​of different bands to a unified dimensional range. For example, if the reflectance value of one band in the original data ranges from 0.1 to 0.9, while another band ranges from 0.3 to 1.2, standardization maps all data to the 0-1 range, facilitating subsequent analysis. For extracting spectral reflectance and difference patterns, a convolutional neural network model is used. Image data is input into the model in time series to capture subtle differences between bands. For instance, if the reflectance value of the red band suddenly increases by 0.2 while the near-infrared band decreases by 0.1 within a certain time period, the model can identify this anomalous combination through multi-layer convolution operations, initially identifying it as a signal of vegetation cover change. When separating anomalous fluctuations, feature selection methods are based on statistical thresholds; for example, fluctuations exceeding twice the average value are marked as anomalous. For example, if the reflectance value of a certain area fluctuates by 0.3 within a certain time period, while the average fluctuation is only 0.05, this area is separated into an anomalous fluctuation set. For the correlation analysis between abnormal fluctuations and potential disasters, fluctuations exceeding a preset threshold of 0.25 are marked as high-risk fluctuations. For example, if a region experiences fluctuations of 0.28, 0.3, and 0.27 for three consecutive days, all exceeding the threshold, it is marked as high-risk. When fusing fluctuation datasets with environmental background data, the data overlay method combines fluctuation data with background information such as soil moisture and rainfall. Assuming the soil moisture data for a high-risk fluctuation area is 80%, far higher than the average of 50%, the fusion suggests the fluctuations may be related to flooding. For the grouping of background fluctuation features, clustering tools can divide features into three groups based on similarity, corresponding to high, medium, and low disaster risks. If a group of features has high fluctuation values ​​and exceeds the soil moisture threshold, it is classified as a high-risk group. When summarizing the distribution of feature combinations, the data integration method aggregates high-risk feature sets from different time periods to form a complete disaster-related feature set. Assuming a region experiences two high-risk fluctuations within three months, the long-term risk trend can be determined after aggregation.

[0059] The disaster risk assessment module is used to obtain a probability assessment value associated with disaster risk based on the combination of potential disaster-related features. If the identified combination of disaster-related features exceeds the preset threshold range, further time series analysis is performed on the abnormal change pattern, and the current change trend is compared with historical data to obtain a probability assessment value associated with disaster risk.

[0060] Furthermore, the disaster risk assessment module includes: a feature analysis and time series analysis unit and a risk probability assessment unit; wherein, the feature analysis and time series analysis unit is used to perform weighted processing on the combination of potential disaster-related features to obtain a weighted feature distribution; the risk probability assessment unit is used to calculate the probability assessment value associated with disaster risk based on the weighted feature distribution, and to determine the potential disaster risk level of the current surface change.

[0061] As a specific implementation of this embodiment, the expression for calculating the disaster risk probability by the risk probability assessment unit is as follows:

[0062]

[0063] In the formula, W(ΔX) is the dynamic weight matrix, CNN(·) is the trained convolutional neural network, P is the disaster risk probability, and X is the original multispectral time series data.

[0064] As a specific implementation of this embodiment, regarding the correlation between disaster characteristics and feature combinations, a data filtering tool is used to initially classify the feature combinations, obtaining a subset of classified features. Based on the classified feature subsets, and considering the characteristics of abnormal changes and change patterns, a data matching method is used to stratify the change patterns, resulting in a set of stratified patterns. Based on the set of stratified patterns, and considering the correspondence between the current trend and the time series, a data mapping tool is used to time-series label the pattern set, determining the labeled time-series feature groups. Based on the labeled time-series feature groups, and combining historical data with data comparison logic, if the change amplitude in the time-series feature group exceeds a preset threshold range, the feature group is weighted to obtain a weighted feature distribution. Based on the weighted feature distribution, and considering the correlation between trend analysis and probability assessment, if the fluctuation frequency in the feature distribution is higher than a preset frequency threshold, a convolutional neural network model is used to perform in-depth analysis of the distribution, obtaining the analyzed risk probability value. Based on the analyzed risk probability value, and combining the correspondence between risk correlation and disaster characteristics, a data fusion method is used to integrate the probability value and disaster characteristics, determining the integrated risk level distribution. Based on the integrated risk level distribution, and considering the mapping relationship between risk associations and change patterns, data archiving tools are used to structurally store the level distribution and determine the final risk pattern archive.

[0065] Furthermore, in studying the correlation between spectral reflectance data and disaster characteristics, the initial classification of feature combinations involves grouping them according to different ranges of reflectance intensity using data filtering tools. For example, assuming a dataset contains reflectance values ​​across multiple spectral bands, features with reflectance values ​​between 0.3 and 0.5 are classified as low-intensity, while those between 0.5 and 0.8 are classified as medium-intensity. When stratifying the characteristics of change patterns, a data matching method is used to stratify patterns according to time span and fluctuation amplitude. For example, in a pattern set, short-term fluctuations with a span of less than 24 hours are classified as transient, while those with a span exceeding 72 hours are classified as long-term. In the time-series labeling of the pattern set, data mapping tools are used to map the pattern set to specific time points. For instance, an abnormal fluctuation is labeled as occurring between 15:00 and 16:00 on a specific date, and combined with environmental background data, it is determined whether it is related to a specific weather event. When weighting the variation amplitude of time-series feature groups, if the fluctuation amplitude of a feature group exceeds a preset threshold of 0.2, its weight can be increased to 1.5 times to highlight its importance. When performing deep analysis using a convolutional neural network model, the fluctuation frequency of the feature distribution is used as input. For example, if the fluctuation frequency of a certain distribution is 3 times per hour, exceeding the preset threshold of 2 times, the model will focus on analyzing its potential risk probability. Regarding the integration of risk probability values ​​and disaster characteristics, a data fusion method is used to correlate features with probability values ​​above 0.75 with historical disaster data to determine if they may correspond to a certain environmental anomaly.

[0066] The disaster risk early warning module includes: a quantitative indicator construction unit, a dynamic early warning threshold setting unit, and an early warning information generation and output unit. The quantitative indicator construction unit is used to construct a quantitative indicator system related to disaster risk based on the probability assessment value. The dynamic early warning threshold setting unit is used to set a dynamically adjustable early warning threshold system for different types of disaster characteristics, combined with application scenarios and risk levels. The early warning information generation and output unit is used to analyze real-time monitoring data based on the early warning threshold system, trigger an early warning signal when the analysis results match an abnormal change pattern, and generate early warning information output including the risk level.

[0067] The closed-loop monitoring module is used to continuously monitor the real-time acquired multispectral data based on the aforementioned early warning reference standard and generate early warning information output including risk levels.

[0068] Raw data streams are acquired from multispectral data sources using real-time acquisition tools. Key features in the data stream are initially extracted to obtain a preliminary feature set. Based on the preliminary feature set, a pre-established early warning standard is used to compare each feature item. If any data in the feature set deviates from the preset reference range, it is marked as a potential anomaly, thus determining an anomaly label set. For the anomaly label set, a change pattern analysis tool is used to perform pattern matching on the data in the label set. If the matching result matches a predefined pattern of abnormal change, a trigger signal is generated, and signal trigger records are obtained. Based on the signal trigger records and the early warning signal generation rules, the anomalies in the records are classified to obtain a classified anomaly level distribution. Based on the classified anomaly level distribution and the risk level classification logic, a data mapping tool is used to map the distribution data to preset risk level intervals to determine the final risk level label. An information output module is used to integrate the final risk level label with the relevant data of abnormal changes to generate structured early warning information content, obtaining a complete early warning information dataset. Through a data transmission tool, the complete early warning information dataset is sent to the relevant monitoring system to complete the information push process and obtain the push completion status.

[0069] Through an early warning information triggering mechanism, high-risk area identification data is obtained from the monitoring system. Using a pre-established regional division logic, the identification data is compared with a geographic information database to obtain the precise distribution of high-risk areas. Based on the precise distribution of high-risk areas, the data acquisition frequency is automatically adjusted. Frequency adjustment commands are issued to image acquisition devices within the area, and the acquisition parameters are updated through the device control module to determine the acquisition plan after frequency adjustment. For the frequency-adjusted acquisition plan, the image acquisition density is increased, and higher-resolution spectral reflectance data is obtained by calling the image sensor interface. The acquired raw images are cleaned using data preprocessing tools to obtain a processed spectral reflectance dataset. Differential data is extracted from the processed spectral reflectance dataset, and a support vector machine algorithm is used for feature classification. If the classification results show that the differential data deviates from a preset threshold range, it is marked as a potential risk point, and a set of marked risk points is obtained. For the set of marked risk points, the trend of risk expansion is analyzed. Using a time series comparison tool, the current set of risk points is matched with historical data. If the matching results show that the distribution range of risk points continues to expand, a risk expansion signal is generated to determine the risk expansion status. Based on the escalating risk, a closed-loop monitoring command is generated and transmitted to the acquisition equipment via a data transmission channel. This automatically adjusts the focus areas and density of subsequent image acquisitions, resulting in an updated monitoring strategy. Using this updated strategy, data acquisition and spectral reflectance difference analysis are continuously performed. The results of each adjustment and analysis are recorded in the system log, obtaining a complete closed-loop monitoring data stream.

[0070] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A multispectral remote sensing disaster monitoring and early warning system utilizing deep learning, characterized in that, include: The spectral data acquisition and preprocessing module is used to acquire surface reflection information from multiple bands, continuously record image data of different time periods based on the surface reflection information to obtain a raw dataset, and preprocess the raw dataset to obtain a clear image dataset. The pattern recognition module is used to perform pattern classification on the clear image dataset to obtain potential disaster-related feature combinations; The disaster risk assessment module is used to perform time series analysis on the combination of potential disaster-related features to obtain a probability assessment value associated with disaster risk; The disaster risk early warning module is used to construct a quantitative indicator system for disaster risk association based on the probability assessment value, and to determine early warning reference standards based on the constructed quantitative indicator system for disaster risk association. The closed-loop monitoring module is used to continuously monitor the real-time acquired multispectral data based on the early warning reference standard and generate early warning information output including risk level. The pattern recognition module includes a spectral feature extraction unit and a pattern classification unit. The spectral feature extraction unit extracts the spectral reflectance characteristics of the clear image dataset, constructs a multi-dimensional feature matrix containing band differences, and analyzes the variation patterns of spectral reflectance differences over different time periods based on the multi-dimensional feature matrix to obtain feature distribution results that reflect surface changes. The pattern classification unit uses a deep learning model to classify spectral reflectance differences based on the feature distribution results, identifying potential disaster-related feature combinations. The pattern classification unit includes: a data standardization and preliminary extraction subunit, an abnormal fluctuation separation and labeling subunit, a feature fusion and clustering subunit, and a disaster feature identification and output subunit. Specifically, the data standardization and preliminary extraction subunit standardizes the spectral reflectance difference data in the feature distribution results to obtain a standardized feature base; based on the standardized feature base, a convolutional neural network model is used to initially extract the spectral reflectance difference patterns to determine the distribution characteristics of the difference patterns. The abnormal fluctuation separation and labeling subunit uses a feature filtering method to separate the abnormal fluctuation portion in the distribution characteristics of the difference patterns to obtain a set of separated abnormal fluctuations; and judges the fluctuation amplitude in the abnormal fluctuation set according to a preset threshold range to obtain a labeled fluctuation dataset. The feature fusion and clustering subunit fuses the labeled fluctuation dataset with environmental background data to determine the fused background fluctuation characteristics; and based on the fused background fluctuation characteristics, a clustering tool is used to group the background fluctuation characteristics to determine the distribution of feature combinations related to disaster risk. The disaster feature identification and output subunit performs statistical analysis on the distribution of feature combinations related to disaster risk and outputs potential disaster-related feature combinations. The disaster risk assessment module includes: a feature analysis and time series analysis unit and a risk probability assessment unit; wherein, the feature analysis and time series analysis unit is used to perform weighted processing on the combination of potential disaster-related features to obtain a weighted feature distribution; the risk probability assessment unit is used to calculate the probability assessment value associated with disaster risk based on the weighted feature distribution, and to determine the potential disaster risk level of the current surface changes; The expression used by the risk probability assessment unit to calculate the disaster risk probability is: ； In the formula, W(ΔX) is the dynamic weight matrix, CNN(·) is the trained convolutional neural network, P is the disaster risk probability, and X is the original multispectral time series data.

2. The multispectral remote sensing disaster monitoring and early warning system utilizing deep learning according to claim 1, characterized in that, The multispectral data acquisition and preprocessing module includes a multispectral data acquisition unit and a data preprocessing unit; The multispectral data acquisition unit is used to acquire surface reflection information from multiple bands and continuously record image data at different time periods to form an original dataset containing dynamic change characteristics. The data preprocessing unit is used to perform denoising and correction operations on each frame of the original dataset using image preprocessing techniques to generate an optimized, clear image dataset.

3. The multispectral remote sensing disaster monitoring and early warning system utilizing deep learning according to claim 2, characterized in that, The data preprocessing unit includes: a first preprocessing subunit, a second preprocessing subunit, a third preprocessing subunit, a fourth preprocessing subunit, and a fifth preprocessing subunit; The first preprocessing subunit is used to perform preliminary denoising on each frame of image to eliminate external interference and obtain the first image dataset after denoising. The second preprocessing subunit is used to perform geometric and radiometric correction on the denoised first image dataset to generate a corrected second image dataset; The third preprocessing subunit uses the comparison threshold range method to perform secondary denoising on the corrected second image dataset to obtain the optimized third image dataset. The fourth preprocessing subunit is used to group the optimized third image dataset according to time order to obtain the grouped fourth image dataset, and extract the image change patterns of the grouped fourth image dataset in each time period to obtain a set of change features. The fifth preprocessing subunit is used to classify outlier segments of the change feature set using the random forest algorithm to obtain an optimized clear image dataset.

4. The multispectral remote sensing disaster monitoring and early warning system utilizing deep learning according to claim 1, characterized in that, The disaster risk early warning module includes: a quantitative indicator construction unit, a dynamic early warning threshold setting unit, and an early warning information generation and output unit; The quantitative indicator construction unit is used to construct a quantitative indicator system for disaster risk association based on the probability assessment value. The dynamic early warning threshold setting unit is used to set a dynamically adjustable early warning threshold system based on different types of disaster characteristics, application scenarios, and risk levels. The early warning information generation and output unit is used to analyze real-time monitoring data based on the early warning threshold system. When the analysis result matches the abnormal change pattern, an early warning signal is triggered, and an early warning information output containing the risk level is generated.

Images