Integrated flood affected area identification method, product, medium and equipment

Abstract

The invention discloses an integrated flood affected area identification method, a product, a medium and equipment, and relates to the technical field of image data processing, and the method comprises the steps: obtaining remote sensing images of a to-be-identified area before and after a disaster occurs; performing calculation based on the remote sensing images before and after the disaster to obtain a plurality of features before and after the disaster; calculating the change rate of the features to obtain a feature change rate image; based on the feature change rate image, performing anomaly scoring by adopting an isolated forest algorithm to obtain an anomaly scoring graph; segmenting the abnormal score map by adopting a threshold segmentation algorithm to obtain a first flood disaster area; and based on the abnormal score map and the first flood disaster area, performing segmentation by adopting an image segmentation model based on self-supervised learning, obtaining a second flood disaster area, and merging the second flood disaster area with the first flood disaster area to obtain a final flood disaster area. The flood disaster area identification precision, efficiency and automation degree can be effectively improved.

Description

Technical Field

[0001] The present application relates to the field of image data processing technology, and in particular to an integrated flood-affected area identification method, product, medium and device. Background Art

[0002] Flood disasters are one of the most destructive natural disasters in the world, with the characteristics of wide impact range, strong suddenness, high frequency and heavy destructiveness. Monitoring and management of flood disasters are crucial to reducing losses, protecting life and property, and promoting regional sustainable development. Accurate extraction of flood inundation areas, that is, accurate identification of flood-affected areas, is the basis of flood disaster management and is crucial to reducing disaster losses. It provides a scientific basis for emergency response and mitigation measures, thereby reducing casualties and economic losses. However, due to the complexity and diversity of flood scenarios, most flood-affected area recognition algorithms are difficult to meet the actual disaster management needs in terms of recognition accuracy, response speed and degree of automation.

[0003] The spectral characteristics, scattering characteristics, water body index, vegetation index, etc. of the flood-affected area will change. The core idea of ​​flood-affected area identification is to detect the changes in the water body area on the two remote sensing images before and after the flood disaster based on the above changes. Therefore, the detection process can be summarized into two parts: difference feature extraction and water body change detection based on difference features. Difference feature extraction refers to the process of identifying or enhancing change features by performing various mathematical operations on remote sensing images before and after the disaster, such as difference, ratio, principal component analysis (PCA), change vector analysis (CVA), etc. Water body change detection based on difference features can be divided into three categories: threshold method, machine learning method and deep learning method. The threshold method classifies image pixels into flood-affected areas and non-flood-affected areas by setting a threshold. The selection of the threshold is highly dependent on the specific data and lacks global applicability. A single threshold detection algorithm is difficult to meet the accuracy and robustness requirements in complex scenarios. Machine learning and deep learning algorithms extract the characteristics of water body changes based on a large amount of sample data and classify these characteristics. Common machine learning algorithms include support vector machines (SVM), random forests (RF), etc. Common deep learning algorithms include convolutional neural networks (CNN), U-Net, etc. These algorithms require a large amount of labeled data and the training process is complex. Compared with machine learning algorithms, deep learning algorithms require larger labeled samples and higher computing resources.

[0004] In actual emergency management, disaster monitoring algorithms usually need to have the characteristics of high accuracy, rapid response and automation. However, the on-site environment after a disaster is complex, and it is difficult to obtain sufficient on-site sample data. At the same time, there are relatively few flood case data, and it is impossible to provide a large number of accurate samples for machine learning or deep learning models to learn and train. This makes it difficult for any of the aforementioned algorithms to simultaneously meet the requirements of high accuracy, rapid response and automation. Summary of the invention

[0005] The purpose of this application is to provide an integrated flood-affected area identification method, product, medium and equipment, which can effectively improve the accuracy, efficiency and automation of flood-affected area identification.

[0006] To achieve the above objectives, this application provides the following solutions:

[0007] In a first aspect, the present application provides an integrated flood-affected area identification method, comprising:

[0008] Obtain remote sensing images of the area to be identified before and after the disaster occurs;

[0009] Based on the remote sensing images before and after the disaster, multiple features before and after the disaster are calculated; the features include: standardized difference water index, standardized difference vegetation index, standardized difference building area index, ratio and dual polarization water index;

[0010] Calculate the rate of change of the feature and obtain a feature change rate image;

[0011] Based on the feature change rate image, the isolation forest algorithm is used to perform anomaly scoring and obtain an anomaly score map;

[0012] The threshold segmentation algorithm is used to segment the abnormal score map to obtain the first flood-affected area;

[0013] Based on the anomaly score map and the first flood-affected area, the image segmentation model based on self-supervised learning is used to segment and obtain the second flood-affected area;

[0014] The first flood-affected area and the second flood-affected area are merged to obtain the final flood-affected area.

[0015] Optionally, the acquiring of remote sensing images of the area to be identified before and after the disaster occurs specifically includes:

[0016] Through the satellite remote sensing data of Sentinel-1 or Sentinel-2 provided by the Google Earth Engine platform, synthetic aperture radar images or optical images of the area to be identified before and after the disaster occur are obtained as remote sensing images of the area to be identified before and after the disaster occur.

[0017] Optionally, the calculation is performed based on the remote sensing images before and after the disaster to obtain multiple features before and after the disaster, specifically including:

[0018] Based on the remote sensing images before and after the disaster, the formula NDWI = (B3-B8) / (B3+B8) was used to calculate the standardized difference water index NDWI before and after the disaster; the formula NDVI = (B4-B8) / (B4+B8) was used to calculate the standardized difference vegetation index NDVI before and after the disaster; the formula NDBI = (B11-B8) / (B11+B8) was used to calculate the standardized difference building area index N before and after the disaster. DBI; the formula Ratio=VV / VH is used to calculate the ratio Ratio before and after the disaster; the formula SDWI=log(10×VV×VH) is used to calculate the dual-polarization water index SDWI before and after the disaster; among them, B3, B4, B8 and B11 represent the satellite remote sensing data of the 3rd, 4th, 8th and 11th bands of Sentinel-2 respectively; VV is the vertical transmission and vertical reception polarization data of Sentinel-1; VH is the vertical transmission and horizontal reception polarization data of Sentinel-1.

[0019] Optionally, by the formula Calculate the rate of change R of feature i; where I after，i and I before，i They are respectively the value of feature i in a pixel of the remote sensing image after the disaster occurs and the value before the disaster occurs.

[0020] Optionally, the abnormality scoring is performed based on the feature change rate image using an isolation forest algorithm to obtain an abnormality scoring map, specifically including:

[0021] The isolation forest algorithm is used to perform anomaly scoring on each pixel of the feature change rate image;

[0022] Based on the anomaly scores of all pixels in the feature change rate image, an anomaly score map of the area to be identified is obtained.

[0023] Optionally, the threshold segmentation algorithm includes Otsu's, Intermodes, Mean, Median, Concavity, Minimum, Triangle, Huang's, Intermeans, ISODATA, MaxEntropy, Li's, Kittler's, Tsai's, Percentile and Zhenzhou's algorithms.

[0024] Optionally, the step of segmenting the first flood-affected area using an image segmentation model based on self-supervised learning based on the anomaly score map and the first flood-affected area to obtain the second flood-affected area specifically includes:

[0025] determining prompt points according to the first flood-affected area and training an image segmentation model based on self-supervised learning based on the prompt points to obtain a trained image segmentation model;

[0026] The anomaly score data in the anomaly score map is reconstructed into three-channel data and input into the trained image segmentation model for segmentation to obtain the second flood-affected area.

[0027] In a second aspect, the present application provides a computer program product, including a computer program, which implements the integrated flood-affected area identification method when executed by a processor.

[0028] In a third aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the integrated flood-affected area identification method.

[0029] In a fourth aspect, the present application provides a computer device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the integrated flood-affected area identification method.

[0030] According to the specific embodiments provided in this application, this application has the following technical effects:

[0031] The present application provides an integrated flood-affected area identification method, product, medium and equipment. By proposing a deep learning segmentation algorithm that integrates an isolation forest algorithm, a threshold segmentation algorithm (TSM) and an image segmentation model based on self-supervised learning, namely, a segmentation everything model (SAM) algorithm, a flood-affected area identification method is used to improve the accuracy and automation level of flood monitoring, thereby effectively improving the accuracy, efficiency and automation level of flood-affected area identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0033] Figure 1 A schematic diagram of the process flow of the integrated flood-affected area identification method provided for this application;

[0034] Figure 2 A technical roadmap for the integrated flood-affected area identification method provided for this application;

[0035] Figure 3 Pre-disaster remote sensing images, post-disaster remote sensing images, feature change rate images and anomaly score maps of plain flood cases provided for this application;

[0036] Figure 4 Pre-disaster remote sensing images, post-disaster remote sensing images, feature change rate images and anomaly score maps of flash flood cases provided for this application;

[0037] Figure 5 Schematic diagram of plain flood identification results provided for this application;

[0038] Figure 6 Schematic diagram of flash flood identification results provided for this application;

[0039] Figure 7 Schematic diagram of the accuracy verification results of mountain torrents and plain floods provided for this application;

[0040] Figure 8 Schematic diagram of grouping and comparative analysis of flood identification results provided for this application;

[0041] Fig. 9 This is a comparison chart of edge details of the TSM algorithm and SAM algorithm results provided in this application. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0043] The emergency management process after a disaster has a high degree of timeliness requirements for disaster detection. Being able to respond to disaster relief in the first place is the key to reducing disaster losses. How to achieve accurate, fast and automated identification of flood-affected areas is a difficult point in the study of flood range extraction. Therefore, this application proposes an integrated flood-affected area identification method that combines the advantages of an isolation forest algorithm, a threshold segmentation algorithm, and a deep learning segmentation algorithm, and names it TSA-SAM. This method can achieve automated flood identification without the need for manual sample selection, and can effectively improve the accuracy, efficiency, and automation of flood-affected area identification.

[0044] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0045] like Figure 1 and Figure 2 As shown, the present application discloses an integrated flood-affected area identification method, comprising:

[0046] Step 1: Obtain remote sensing images of the area to be identified before and after the disaster.

[0047] This application mainly uses the satellite remote sensing data of Sentinel-1 or Sentinel-2 provided by the Google Earth Engine platform to obtain synthetic aperture radar images or optical images of the area to be identified before (pre-disaster) and after (post-disaster) the disaster occurs as remote sensing images of the area to be identified before and after the disaster occurs.

[0048] Specifically, for the selection of the area to be identified, as a specific embodiment, in order to fully test the applicability of the TSA-SAM method of this application in various terrains and environments, this application selected 10 flood cases distributed in 8 regions at different latitudes and altitudes for algorithm testing. Among them, the 5 plain flood cases are case (a) occurring in region A, case (b) occurring in region B, case (c) occurring in region C, case (d) occurring in region D, and case (e) occurring in region E. The 5 mountain torrent cases are case (f) occurring in region F, case (g) occurring in region G, and case (h), case (i), and case (j) occurring in region H. The detailed information of each flood case is shown in Table 1.

[0049] Table 1 Detailed information of each flood case

[0050]

[0051] As a specific embodiment, the data used in this application is the satellite remote sensing data of Sentinel-1 and Sentinel-2 provided by the public data archive of the Google Earth Engine (GEE) platform for obtaining the data source of remote sensing images before and after the disaster in the area to be identified. Sentinel-1 provides Synthetic Aperture Radar (SAR) images, which have the advantages of being able to penetrate clouds and not being affected by weather, and are suitable for flood monitoring under rainy or typhoon weather conditions. However, there is speckle noise in the SAR images of Sentinel-1. The formation of speckle is caused by the randomness of the position of the scatterer, which causes the received signal to lose consistency in phase. Although speckle filtering technology can reduce noise, it often sacrifices the detailed features in the image. Especially in mountainous environments, the speckle problem makes it difficult to accurately identify flood-affected areas. Sentinel-2 provides high-resolution optical images, including multiple bands from visible light to near-infrared and short-wave infrared. It can provide rich spectral information of surface objects, has advantages in identifying surface vegetation, water bodies and soil, and is not affected by speckle noise. However, the limitation of optical remote sensing data is that it is easily affected by weather, especially when floods are often accompanied by long-term rainy or typhoon weather, the obstruction of clouds will seriously affect the acquisition and use of flood identification data. Therefore, this application selects different satellites to obtain corresponding data sets according to the data requirements in different flood scenarios. In the identification of flood-stricken areas in mountainous areas, this application gives priority to the use of Sentinel-2 optical images as the data source to avoid the influence of speckle noise in mountainous areas. In plain areas, this application selects Sentinel-1 SAR images as the data source to ensure effective flood identification even in severe weather conditions.

[0052] The Sentinel-1 data provided on the GEE platform is a dual-polarization C-band synthetic aperture radar data with an operating frequency of 5.405GHz. It has been calibrated using the Sentinel-1 toolbox and can be directly obtained through the link ee.ImageCollection("COPERNICUS / S1_GRD"). This application also utilizes the vertical transmission and vertical reception (VV) and vertical transmission and horizontal reception (VH) polarization band data of Sentinel-1. The Sentinel-2 data provided on the GEE platform is a multi-spectrometer surface reflectivity data that has been corrected for radiation and can be accessed through the link ee.ImageCollection("COPERNICUS / S2_SR_HARMONIZED"). This application uses bands 2 to 8A and band 11 in the Sentinel-2 data. In order to minimize the impact of cloud cover, this application selects remote sensing image data with less than 20% cloud cover within a certain period of time before and after the disaster for subsequent preprocessing and analysis. If there is no qualified data within one month before the disaster, the data of the same period of the previous year will be used as the pre-disaster image. The specific data corresponding to the 10 flood cases and the data acquisition date are shown in Table 2.

[0053] Table 2 Data information of each flood case

[0054]

[0055]

[0056] After acquiring the remote sensing image data, subsequent calculations are performed based on the remote sensing images, such as Figure 1 and Figure 2 As shown in FIG. 1 , the TSA-SAM method of the present application specifically includes the following key steps: 1) Feature calculation: Calculate the water index, vegetation index and other features of the remote sensing image before and after the disaster, as shown in step 2. 2) Feature change rate image calculation: Calculate the change rate of the above features before and after the disaster, as shown in step 3. 3) Anomaly score map: Take the feature change rate image as input, use the isolation forest algorithm to calculate the normalized anomaly score (anomaly score) of each pixel in the feature change rate image, and generate an anomaly score map, as shown in step 4. 4) Threshold segmentation: Use a variety of unsupervised threshold segmentation algorithms to automatically segment the anomaly score map to obtain the first flood-affected area, as shown in step 5. 5) SAM segmentation: Automatically select prompt points based on the threshold segmentation result (the first flood-affected area), and apply the SAM model for segmentation to obtain the second flood-affected area R2, as shown in step 6. 6) Merge the segmentation results: Merge R1 and R2 to obtain the final flood-affected area, as shown in step 7.

[0057] Step 2: Calculate based on remote sensing images before and after the disaster to obtain multiple features before and after the disaster.

[0058] In this application, both Sentinel 1 and Sentinel 2 data are preprocessed through the GEE platform. For Sentinel 2 data, in order to minimize the impact of cloud cover and improve data availability, the quality assessment band QA60 is first used to remove clouds and cirrus cover in each image. Then, all bands are resampled to 10-meter resolution, and the ImageCollection.median function provided by the GEE platform is used to calculate the median image of each band before and after the disaster. Taking the pre-disaster data of the Sentinel 2 blue band (Band 2) as an example, the specific steps are as follows: First, the images of Sentinel 2 Band 2 are superimposed to form a time-series composite image of Band 2. For each pixel p on the composite image, the average of all values ​​of the pixel is calculated, and the pixel value closest to the average value in the Band 2 reflectance of pixel p in the original composite image is assigned to pixel p. Then, the above operation is performed pixel by pixel to obtain the median image of Band 2. The above operation is performed on each band before and after the disaster to obtain a synthetic median map before and after the disaster.

[0059] Afterwards, based on the synthesized Sentinel-2 data, i.e. the preprocessed Sentinel-2 data, the Standardized Difference Water Index (NDWI) was calculated to highlight the water body information, the Standardized Difference Vegetation Index (NDVI) was calculated to highlight the vegetation information, and the Standardized Difference Building Area Index (NDBI) was calculated to enhance the building and bare soil information. Based on the preprocessed VV and VH polarization data of Sentinel-1, the ratio of VV to VH was calculated as one of the features. This ratio helps to enhance the contrast effect of data from different polarization channels. In addition, the Sentinel-1 Dual-Polarized Water Index (SDWI) was calculated. This index is designed to improve the accuracy of water body monitoring, can effectively distinguish water bodies from other surface types, and enhance the ability to identify floods. The above indices are all the features to be calculated, and the specific calculation formulas of each feature are shown in Table 3.

[0060] Table 3 Characteristic calculation formula

[0061] feature Calculation formula NDWI (B3-B8) / (B3+B8) NDVI (B4-B8) / (B4+B8) NDBI (B11-B8) / (B11+B8) Ratio VV / VH SDWI log(10×VV×VH)

[0062] Among them, B3, B4, B8 and B11 represent the satellite remote sensing data of the 3rd, 4th, 8th and 11th bands of Sentinel-2 respectively; VV is the vertical transmission and vertical reception polarization data of Sentinel-1; VH is the vertical transmission and horizontal reception polarization data of Sentinel-1.

[0063] Step 3: Calculate the rate of change of the feature and obtain the feature change rate image.

[0064] The following formula is used to calculate the rate of change R of the characteristics before and after the disaster:

[0065]

[0066] Among them, I after，i and I before，i are the values ​​of feature i in a pixel of a remote sensing image after the disaster and before the disaster, respectively. The rate of change helps to eliminate the changes caused by differences in brightness or exposure between different images, highlight the real change information, suppress the change information of the background, reduce environmental influences such as image illumination change and shadow change, and improve the robustness of the results. The feature change rate of each feature in step 2 is calculated pixel by pixel, and a new feature change rate image is formed based on the feature change rate corresponding to each pixel. At the same time, since the rate of change takes into account the intensity of the initial state, the result is a relative change value, which can make the image results in different regions or at different times more comparable.

[0067] Step 4: Based on the feature change rate image, the isolation forest algorithm is used to perform anomaly scoring and obtain an anomaly score map.

[0068] The Isolation Forest Algorithm was proposed by Liu Fei, Zhou Zhihua and others in 2008. It is an unsupervised anomaly detection method based on the decision tree algorithm. The algorithm recursively constructs a binary tree to randomly partition the data set until all sample points are isolated, and characterizes the so-called degree of anomaly based on the path length required for the sample point to be isolated. The path length is the number of edges from the root node to the leaf node. Abnormal samples are defined as "outliers that are easily isolated" and can be isolated through a shorter path. The algorithm does not use indicators such as distance and density to describe the difference between samples and other samples, nor does it specify the range or pattern of normal samples. It can detect anomalies very simply and efficiently, and is widely used in the industry. The Isolation Forest Algorithm included in the Scikit-learn toolkit takes each pixel in the input image as a sample and calculates the normalized anomaly score pixel by pixel.

[0069] In this application, an isolation forest algorithm is used to perform an abnormality score on each pixel in the feature change rate image, and an abnormality score map of the area to be identified is obtained based on the abnormality scores of all pixels in the feature change rate image. Specifically, the abnormality score is represented by a numerical value in the range of [-1,1]. The closer the abnormality score is to 1, the lower the abnormality of the pixel is, and the lower the possibility that it is a flood pixel is; on the contrary, the closer the abnormality score is to -1, the higher the abnormality of the pixel is, and the higher the possibility that it is a flood pixel is.

[0070] Step 5: Use the threshold segmentation algorithm to segment the anomaly score map to obtain the first flood-affected area.

[0071] This application uses a total of 16 unsupervised threshold segmentation algorithms to segment the abnormal score map, divide the abnormal score map into two categories pixel by pixel, and calculate the mean of the abnormal score corresponding to each pixel in the two categories. Among the 16 methods, Otsu's, Intermodes, Mean, Median, Concavity, Minimum, Triangle and Percentile belong to histogram-based threshold segmentation algorithms. This type of method is widely used in image segmentation with significant differences between background and foreground due to its direct and fast characteristics. Huang's, Intermeans, ISODATA, etc. are methods based on iterative calculation and cluster analysis. These methods improve the accuracy of segmentation through a continuously refined iterative calculation process and can effectively process images with complex structures or more noise. MaxEntropy, Li's, Kittler's, and Tsai's are methods based on information theory. These methods optimize threshold selection by maximizing information retention or minimizing error functions, and are suitable for processing complex images. Zhenzhou's method smoothes the histogram through Fourier transform and finds key points to determine the threshold. Next, the category with low mean anomaly score (i.e., high degree of anomaly) was judged as flood (1), and the category with high mean anomaly score (i.e., low degree of anomaly) was judged as non-flood (0), thus generating 16 different classification results. Finally, these 16 results were superimposed. When a pixel value was ≥15, that is, when at least 15 algorithms judged the pixel to be flood, the pixel would be marked as flood in the final threshold segmentation result. The first flood-affected area thus obtained was recorded as R1.

[0072] Step 6: Based on the anomaly score map and the first flood-affected area, the image segmentation model based on self-supervised learning is used to perform segmentation to obtain the second flood-affected area.

[0073] SAM is an image segmentation model based on self-supervised learning developed by Meta Lab. It can adapt to various segmentation tasks through simple user interaction (such as clicking or drawing polygons) prompt engineering (prompt), thereby improving the flexibility and accuracy of processing complex visual tasks, and can generate multiple effective masks in the face of uncertainty. The model architecture of SAM includes an image encoder, a prompt encoder and a mask decoder, which respectively generate a one-time image embedding, encode the prompt into an embedding vector in real time, and predict the segmentation mask. These components work together to efficiently generate accurate segmentation results. The SAM model has strong adaptability and flexibility, and can still generate high-quality segmentation results in the absence of precise annotations. As a specific embodiment, the present application adopts the SAM-HQ algorithm, which uses a more advanced training model than the initial SAM, and has higher segmentation accuracy and processing capabilities for complex boundary objects. Its advantage is that objects with fuzzy boundaries can be segmented more accurately, significantly improving the overall segmentation effect.

[0074] It should be noted that deep learning image processing models are often designed and pre-trained based on natural images of three color channels, RGB (red, green, and blue). In this application, the input data of the SAM-HQ model is the abnormal score map output by the isolation forest algorithm, and the output result of the isolation forest algorithm is single-channel data. Therefore, by copying the single-channel data to form three-channel data, these results can be accepted and processed by SAM.

[0075] In addition, unlike traditional supervised classification methods, the prompt points required by the SAM model are not sample points in traditional machine learning supervised classification, but provide a preliminary area of ​​interest so that the model can perform detailed feature extraction and boundary recognition around these points. The original intention of the design of this method is to determine the final segmentation result by comprehensively analyzing the global features and local details of the entire image, rather than relying entirely on the prompt points, thereby avoiding the limitation of over-reliance on a single prompt information, while ensuring the flexibility and robustness of the segmentation results. This application randomly selects prompt points based on the result R1 of the threshold segmentation. First, the total number of prompt points is set to 40; then, in order to ensure the balance of positive and negative samples, 15 pixels are randomly selected from the flood (1) and non-flood (0) pixels in R1 as positive and negative prompt points; finally, for the remaining 10 prompt points, the number of positive and negative prompt points is determined based on the ratio of flood and non-flood pixels in the result R1. Specifically, let the number of flood pixels in the result R1 be N pos , the number of non-flood pixels is N neg , define the ratio r as:

[0076]

[0077] In flash flood scenes, the flood areas (positive class) are usually concentrated, and r is often less than 0.5. It is necessary to give priority to enhancing the positive class information to ensure the correct recognition rate of the positive class and minimize the false alarm rate of the negative class (non-flood area). Therefore, it is necessary to give priority to allocating a larger proportion of samples to the positive class prompt points. On the contrary, in plain floods, the flood areas (positive class) are usually widely distributed, and r is often greater than 0.5. It is necessary to give priority to enhancing the negative class (non-flood area) information to improve the correct recognition rate of the negative class and reduce the false alarm rate of the positive class. Considering the needs of the above two scenarios, the number of prompt points allocated to the positive class (flood area) of 10 prompt points is set to 10-(10×r), and the number of prompt points allocated to the negative class (non-flood area) is set to 10×r. This method realizes the flexible adjustment of the sample ratio under the premise of controlling the total number of samples, which can effectively optimize the performance based on image segmentation.

[0078] The prompt points selected in the above steps are used to train the SAM-HQ segmentation model to obtain a trained image segmentation model. Subsequently, the trained image segmentation model is used to reconstruct the three-channel data obtained by the anomaly score data in the anomaly score map generated by the isolation forest algorithm to obtain the flood distribution range based on the SAM method. In order to improve the stability of the segmentation results, 20 repeated segmentation operations are performed, and the 20 segmentation results are superimposed. When a pixel value is ≥18, that is, at least 18 times SAM determines that the pixel is a flood, the pixel will be marked as a flood in the final SAM segmentation result. The second flood-affected area obtained is recorded as R2.

[0079] Step 7: Merge the first flood-affected area and the second flood-affected area to obtain the final flood-affected area.

[0080] Compared with the dynamic threshold method, the SAM model has obvious advantages in maintaining segmentation consistency and integrity, and can avoid the uncertainty and fragmentation problems caused by the selection of a single threshold. However, the SAM model has some shortcomings when dealing with small boundary changes. In order to make full use of the respective advantages of the two algorithms, this application integrates the segmentation results (R2) based on the SAM model with the segmentation results (R1) of the dynamic threshold method to obtain the final flood-affected area, denoted as R3. Specifically, the segmentation results of R1 and R2 are first converted into independent vector objects, and each connected pixel area is regarded as an independent object. Then, the intersection of the segmentation results of R1 and R2 is extracted. For these intersection areas, the corresponding objects in R1 are retained as the final flood-affected area R3. By merging the results of the two segmentation algorithms, not only the advantages of the SAM algorithm in maintaining the consistency and integrity of image segmentation are brought into play, but also the sensitivity of the dynamic threshold method in processing detail edges is fully utilized, thereby achieving a more accurate and stable identification of the flood-affected area.

[0081] In addition, based on the TSA-SAM method provided in this application, it is also necessary to analyze the remote sensing images before and after the disaster through manual visual interpretation, and accurately digitize the range of water body changes before and after the flood, so as to verify the accuracy of the flood-affected range extracted by the TSA-SAM method of this application.

[0082] Taking the flood range obtained by manual visual interpretation and digitization as a reference, the four indicators of Overall Accuracy (OA), Precision (P), Recall (R) and F1-Score (F1) are used to evaluate the flood recognition results obtained by the threshold segmentation algorithm, SAM algorithm and the TSA-SAM method of this application. The evaluation indicators are shown in Table 4.

[0083] Table 4 Evaluation indicators

[0084]

[0085] Among them, TP means that the classifier predicts that the sample is a positive sample, and it is actually a positive sample, that is, the number of positive samples correctly identified; FP means that the classifier predicts that the sample is a positive sample, but it is actually a negative sample, that is, the number of false positive negative samples; TN means that the classifier predicts that the sample is a negative sample, and it is actually a negative sample, that is, the number of negative samples correctly identified; FN means that the classifier predicts that the sample is a negative sample, but it is actually a positive sample, that is, the number of missed positive samples.

[0086] This application uses remote sensing data before and after the disaster to calculate the feature change rate image of each feature, and calculates the anomaly score based on the isolation forest algorithm. Figure 3 and Figure 4The images before and after the flood disaster, the characteristic change rate images and the anomaly score map containing the isolated forest anomaly score data of the 10 case areas in this application are shown, among which: Figure 3 (a1), (b1), (c1), (d1) and (e1) are the pre-disaster remote sensing images of Case (a), Case (b), Case (c), Case (d) and Case (e), respectively; Figure 3 (a2), (b2), (c2), (d2) and (e2) are the post-disaster remote sensing images of Case (a), Case (b), Case (c), Case (d) and Case (e), respectively; Figure 3 (a3), (b3), (c3), (d3) and (e3) are the feature change rate images of case (a), case (b), case (c), case (d) and case (e), respectively; Figure 3 (a4), (b4), (c4), (d4) and (e4) are the anomaly score graphs of case (a), case (b), case (c), case (d) and case (e), respectively. Figure 4 (f1), (g1), (h1), (i1) and (j1) are the pre-disaster remote sensing images of Case (f), Case (g), Case (h), Case (i) and Case (j), respectively; Figure 4 (f2), (g2), (h2), (i2) and (j2) are the post-disaster remote sensing images of Case (f), Case (g), Case (h), Case (i) and Case (j), respectively; Figure 4 (f3), (g3), (h3), (i3) and (j3) are the characteristic change rate images of case (f), case (g), case (h), case (i) and case (j), respectively; Figure 4 (f4), (g4), (h4), (i4) and (j4) are the anomaly score maps of case (f), case (g), case (h), case (i) and case (j) respectively. It can be clearly seen from the figure that the isolation forest algorithm effectively highlights the scope of the flood area. The high anomaly area (anomaly score close to -1) obviously corresponds to the flood range with abnormal changes, while the low anomaly area (anomaly score close to 1) corresponds to the non-flood area.

[0087] This application uses three methods, namely, SAM algorithm, threshold segmentation algorithm and TSA-SAM method, to automatically identify floods in two scenarios, plain flood and mountain flood. The identification results are as follows: Figure 5 and Figure 6 As shown. Among them, Figure 5 (a1), (b1), (c1), (d1) and (e1) are the rate of change images and visually interpreted flood extents of Case (a), Case (b), Case (c), Case (d) and Case (e), respectively; Figure 5(a2), (b2), (c2), (d2) and (e2) are the TSM segmentation results R1 of case (a), case (b), case (c), case (d) and case (e) respectively; Figure 5 (a3), (b3), (c3), (d3) and (e3) are the SAM segmentation results R2 of case (a), case (b), case (c), case (d) and case (e) respectively; Figure 5 (a4), (b4), (c4), (d4) and (e4) are R3 obtained by the TSM-SAM method of this application for case (a), case (b), case (c), case (d) and case (e), respectively. Figure 6 (f1), (g1), (h1), (i1) and (j1) are the rate of change images and visually interpreted flood extents of Case (f), Case (g), Case (h), Case (i) and Case (j), respectively; Figure 6 (f2), (g2), (h2), (i2) and (j2) are the TSM segmentation results R1 of case (f), case (g), case (h), case (i) and case (j), respectively; Figure 6 (f3), (g3), (h3), (i3) and (j3) are the SAM segmentation results R2 of case (f), case (g), case (h), case (i) and case (j), respectively; Figure 6 (f4), (g4), (h4), (i4) and (j4) are R3 obtained by the TSM-SAM method of this application for case (f), case (g), case (h), case (i) and case (j), respectively.

[0088] Figure 7The flood identification accuracy evaluation results for two scenarios, flash flood and plain flood, are shown. From the numerical results of the evaluation indicators, the TSA-SAM method shows the best classification effect in both flash flood and plain flood scenarios. In the flash flood scenario, the average overall accuracy of the TSA-SAM method is as high as 0.9823, and the average F1-Score is 0.8223, which are significantly better than the methods using SAM or threshold segmentation alone. Especially in the H area, the overall accuracy and F1-Score of the method reached 0.9970 and 0.9567 respectively. In addition, the performance of the method in other areas such as G and F is also far better than that of a single method. This shows that when dealing with flash flood scenarios with complex terrain and changeable climate, the integrated algorithm can effectively improve the accuracy of flood detection. In the plain flood case, the TSA-SAM method still performs best, with an average overall accuracy of about 0.9460 and an average F1-Score of 0.8701. Especially in area D, its overall accuracy reaches 0.9779 and F1-Score reaches 0.9022. This advantage is also obvious in other areas such as areas B and C, with F1-Score reaching 0.9142 and 0.8445 respectively. Therefore, the TSA-SAM method also shows excellent accuracy and adaptability in plain flood scenarios.

[0089] In summary, TSM-SAM's recognition accuracy in 10 cases of mountain torrents and plain floods is significantly better than that of a single algorithm, with an average overall accuracy of 0.9823 and 0.9460, and an average F1-Score of 0.8223 and 0.8701, respectively. TSM-SAM achieves automatic, highly robust, and highly adaptable flood-affected area identification, has high practical value, and provides strong decision-making support for disaster emergency management.

[0090] like Figure 8As shown, this application calculates the mean, median and upper and lower quartile range of the above evaluation indicators in groups from two dimensions: scene (mountain torrents and plain floods) and algorithm (TSM, SAM, TSM-SAM). From the perspective of different scenes, the median distribution of the precision indicators in mountain torrents and plain floods is roughly similar, all concentrated between 0.7 and 0.9, indicating that the classification accuracy of the three algorithms in these two scenes is generally consistent. However, there are more outliers in the box plot of the plain flood scene, and the data is more discrete, indicating that there are greater differences in the identification of plain floods among the three algorithms. From the perspective of different algorithms, the median and mean of the SAM method are relatively high, indicating that it may have better performance in some cases, but the data is more discrete, and a large number of outliers show that it is not stable enough under different conditions. The evaluation indicators of the TSM method show high robustness, but the mean of Precision is significantly lower, indicating that the TSM method often misjudges non-flood areas (negative class) as flood areas (positive class), resulting in more false positives. The mean and median of the accuracy evaluation results of the TSM-SAM integrated algorithm are the highest, and the distribution is the most concentrated with small variability, which means that while ensuring high robustness, it can better capture the real flood area.

[0091] In summary, whether in mountain flood or plain flood scenarios, the TSA-SAM method combines the high sensitivity of SAM with the robustness of TSM, and demonstrates robustness, consistency, accuracy, and generalization capabilities that are superior to those of a single method. It significantly improves the reliability and practicality of the flood monitoring automation process and greatly enhances the practical value of actual business processes.

[0092] In addition, this application also studies the adaptation of the isolation forest algorithm to the deep learning framework. In this application, the isolation forest algorithm demonstrates its unique advantage in highlighting abnormal information in disaster monitoring applications. The isolation forest algorithm provides a clear classification of targets and backgrounds for subsequent deep learning models by locating areas of abnormal high or low values, enabling them to extract specific information more effectively. The single-channel data output by the isolation forest algorithm can be simply converted to three-channel data, so that the key information in the remote sensing data can be accepted and processed by the standard deep learning architecture. This conversion process does not sacrifice the key information of the data, but can efficiently pass important abnormal features to the deep learning model. In practical applications, remote sensing data often comes from different types of sensors, such as optical images, SAR images, thermal infrared images, etc. There are significant differences between these data in terms of spectrum, resolution, imaging mechanism, etc. The pre-processing of the isolation forest algorithm allows these heterogeneous image data to be uniformly input into the deep learning model for processing, reducing the dependence on a single data source or image type, so that different types of image data can be directly used for anomaly detection and segmentation processing, making it easier to integrate multi-source heterogeneous data into a unified detection framework, greatly enhancing the model's adaptability to different scenarios and different data sources, and making the unified detection model widely used in more complex and diverse disaster monitoring environments. In terms of extension, the isolation forest algorithm can form more combinations with other deep learning algorithms such as classification and segmentation to build more complex integrated detection frameworks. In summary, the isolation forest algorithm not only improves the accuracy and practicality of the overall algorithm, but also provides new ideas for the application of deep learning models in the field of disaster detection to a certain extent.

[0093] For the advantages of combining deep learning segmentation algorithm and traditional threshold segmentation algorithm, Fig. 9 The comparison shows the details of the segmentation results of the SAM algorithm and the threshold segmentation algorithm, where: Fig. 9 (a1) in the figure is the TSM segmentation details of case (a). Fig. 9 (a2) in the figure is the SAM segmentation details of case (a). Fig. 9 (b1) in the figure is the TSM segmentation details of case (b). Fig. 9(b2) in the figure is the SAM segmentation details of case (b). The threshold segmentation algorithm can achieve pixel-level edge processing, and the edge performance is very good, but it is also prone to errors such as fragmentation, resulting in poor coherence and integrity. In contrast, the SAM algorithm can accurately identify the main position of the flood range and has good coherence for the identified objects, but its edge details are fuzzy and smooth, resulting in the loss of edge details. This may be because the SAM algorithm is a pre-trained model based on a general data set, and it is not specifically optimized for remote sensing data during training. When processing remote sensing images, the model may not be able to accurately capture the specific boundary features of the remote sensing images, resulting in poor performance of the model in the edge areas of remote sensing data. In addition, when extracting features, deep learning models often emphasize the importance of global contextual information, especially in higher-level convolutional layers, where the model tends to capture large-scale semantic information. However, this enhancement of global information is often at the expense of local edge information. In edge areas, especially those adjacent to the background or other object boundaries, the model may not be able to effectively distinguish subtle boundary differences, resulting in inaccurate edge detection.

[0094] The TSM-SMA flood identification method proposed in this application combines the respective advantages of deep learning algorithms and traditional threshold segmentation algorithms. The SAM model represents the forefront of current deep learning image segmentation algorithms, and it has significant advantages in feature extraction and processing of complex scenes. Traditional threshold segmentation methods do not rely on large-scale data sets for training, and the calculation process is relatively simple and fast, which is particularly suitable for real-time applications and application scenarios with limited computing resources. The TSM-SMA algorithm can make full use of the powerful feature extraction and complex scene processing capabilities of deep learning, while inheriting the efficiency and robustness of traditional threshold methods. Combining deep learning with traditional threshold segmentation, machine learning and other methods is an effective strategy to improve the accuracy and efficiency of business processes. It is an algorithm strategy that should be given priority in scenarios with limited computing resources and scenarios with high requirements for algorithm lightness and timeliness.

[0095] In summary, this application develops and verifies an integrated flood-affected area identification method, which comprehensively applies the isolation forest algorithm, multiple threshold segmentation algorithms and the SAM segmentation model, and uses remote sensing images before and after the disaster to achieve unsupervised, rapid response and high-precision flood-affected area identification for various flood scenarios. The above verification results show that: 1) The isolation forest algorithm effectively highlights the flood-affected areas, enhances the adaptability of the method to different scenarios and different data sources, and improves the accuracy and practicality of the overall algorithm; 2) The TSM-SAM method has achieved high flood recognition accuracy in ten cases of mountain torrents and plain floods. The average overall accuracy of recognition in mountain torrents and plain floods is 0.9823 and 0.9460 respectively, and the F1-Score average reaches 0.8223 and 0.8701 respectively; 3) Compared with the single threshold segmentation algorithm and deep learning segmentation algorithm, the TSM-SAM method has shown significant performance improvement, and can fully utilize the powerful feature extraction and complex scene processing capabilities of deep learning and the efficiency and robustness of traditional methods at the same time, and realize lightweight real-time applications for scenarios with limited computing resources. Therefore, the TSM-SAM method proposed in this application can realize automatic and rapid flood recognition in various scenarios, improve the accuracy of flood recognition under different scenarios and data source conditions, and has practical application value in complex and diverse remote sensing disaster monitoring tasks, and can provide a decision-making basis for the formulation of timely and effective disaster management measures.

[0096] In some embodiments, the present application also provides a computer program product, including a computer program, which implements the integrated flood-affected area identification method when executed by a processor.

[0097] In some embodiments, the present application also provides a computer-readable storage medium having a computer program stored thereon, which implements the integrated flood-affected area identification method when executed by a processor.

[0098] In some embodiments, the present application also provides a computer device, including a processor, a memory, an input / output interface (I / O for short), a communication interface, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the integrated flood-affected area identification method.

[0099] The processor, memory and input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store pending transactions. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the integrated flood-affected area identification method is implemented.

[0100] It should be noted that the object information (including but not limited to object device information, object personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the object or fully authorized by all parties, and the collection, use and processing of relevant data need to comply with relevant laws, regulations and standards of relevant countries and regions.

[0101] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ReadOnlyMemory, ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0102] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0103] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. An integrated flood-affected area identification method, characterized in that: include: Obtain remote sensing images of the area to be identified before and after the disaster occurs; Based on the remote sensing images before and after the disaster, multiple features before and after the disaster are obtained; The characteristics include: standardized difference water index, standardized difference vegetation index, standardized difference building area index, ratio and dual polarization water index; Calculate the rate of change of the feature and obtain a feature change rate image; Based on the feature change rate image, the isolation forest algorithm is used to perform anomaly scoring and obtain an anomaly score map; The threshold segmentation algorithm is used to segment the abnormal score map to obtain the first flood-affected area; Based on the anomaly score map and the first flood-affected area, the image segmentation model based on self-supervised learning is used to segment and obtain the second flood-affected area; The first flood-affected area and the second flood-affected area are merged to obtain the final flood-affected area.

2. The integrated flood-affected area identification method according to claim 1, characterized in that: The obtaining of remote sensing images of the area to be identified before and after the disaster occurs specifically includes: Through the satellite remote sensing data of Sentinel-1 or Sentinel-2 provided by the Google Earth Engine platform, synthetic aperture radar images or optical images of the area to be identified before and after the disaster occur are obtained as remote sensing images of the area to be identified before and after the disaster occur.

3. The integrated flood-affected area identification method according to claim 2, characterized in that: The calculation is performed based on the remote sensing images before and after the disaster to obtain multiple features before and after the disaster, including: Based on the remote sensing images before and after the disaster, the formula NDWI = (B3-B8) / (B3+B8) was used to calculate the standardized difference water index NDWI before and after the disaster; the formula NDVI = (B4-B8) / (B4+B8) was used to calculate the standardized difference vegetation index NDVI before and after the disaster; the formula NDBI = (B11-B8) / (B11+B8) was used to calculate the standardized difference building area index N before and after the disaster. DBI; the formula Ratio=VV / VH is used to calculate the ratio Ratio before and after the disaster; the formula SDWI=log(10×VV×VH) is used to calculate the dual-polarization water index SDWI before and after the disaster; among them, B3, B4, B8 and B11 represent the satellite remote sensing data of the 3rd, 4th, 8th and 11th bands of Sentinel-2 respectively; VV is the vertical transmission and vertical reception polarization data of Sentinel-1; VH is the vertical transmission and horizontal reception polarization data of Sentinel-1.

4. The integrated flood-affected area identification method according to claim 3, characterized in that: By formula Calculate the rate of change R of feature i; where I after，i and I before，i They are respectively the value of feature i in a pixel of the remote sensing image after the disaster occurs and the value before the disaster occurs.

5. The integrated flood-affected area identification method according to claim 4, characterized in that: Based on the feature change rate image, the isolation forest algorithm is used to perform anomaly scoring to obtain an anomaly scoring map, which specifically includes: The isolation forest algorithm is used to perform anomaly scoring on each pixel of the feature change rate image; Based on the anomaly scores of all pixels in the feature change rate image, an anomaly score map of the area to be identified is obtained.

6. The integrated flood-affected area identification method according to claim 5, characterized in that: The threshold segmentation algorithms include Otsu's, Intermodes, Mean, Median, Concavity, Minimum, Triangle, Huang's, Intermeans, ISODATA, MaxEntropy, Li's, Kittler's, Tsai's, Percentile and Zhenzhou's algorithms.

7. The integrated flood-affected area identification method according to claim 5, characterized in that: The method uses an image segmentation model based on self-supervised learning to perform segmentation based on the abnormal score map and the first flood-affected area to obtain the second flood-affected area, specifically comprising: determining prompt points according to the first flood-affected area and training an image segmentation model based on self-supervised learning based on the prompt points to obtain a trained image segmentation model; The anomaly score data in the anomaly score map is reconstructed into three-channel data and input into the trained image segmentation model for segmentation to obtain the second flood-affected area.

8. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the integrated flood-affected area identification method according to any one of claims 1 to 7 is implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the integrated flood-affected area identification method according to any one of claims 1 to 7 is implemented.

10. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the integrated flood-affected area identification method according to any one of claims 1 to 7.

Images