Harmonic spectrum information cutting-based flood inundation range remote sensing rapid extraction method

Abstract

The invention discloses a harmonic spectrum information cutting-based flood inundation range remote sensing rapid extraction method, which comprises the steps of S1, converting an obtained remote sensing image from an RGB color space to an HSV color space, extracting spectral information of the image, and decomposing the spectral information into harmonic spectrum components; s2, constructing an improved clustering model based on a density peak value clustering algorithm, and clustering pixel points in the remote sensing image into a suspected water body class and a non-water body class; s3, optimizing a clustering result through a semantic segmentation network, taking the clustered image as input, and outputting a water body and non-water body segmentation result; and S4, in combination with geographic coordinate information of the remote sensing image, geocoding is performed on the segmentation result, and a thematic map containing the flood inundation range is generated. According to the method, the accuracy and efficiency of flood inundation range extraction are improved, and accurate distinguishing between a water body and a non-water body and rapid obtaining of inundation range changes are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of water conservancy remote sensing images, and in particular to a method for rapid remote sensing extraction of flood inundation range based on harmonic spectrum information cutting. Background Art

[0002] Flood disasters are characterized by their sudden onset and widespread impact. Rapid and accurate information on the inundation area and its dynamic changes is a core requirement for emergency relief, disaster assessment, and flood control planning. Remote sensing technology, with its ability to simultaneously observe large areas, has become a crucial tool for flood monitoring. However, the spectral characteristics of water and non-water bodies (such as shadows, bare land, and vegetation) in flood scenes are susceptible to interference from factors such as weather conditions, land cover type, and sediment content, making it difficult for traditional methods to achieve high-precision segmentation.

[0003] Traditional methods for extracting flood inundation areas are primarily based on spectral thresholding and traditional machine learning algorithms. Spectral thresholding segments water bodies by setting a fixed spectral threshold (such as the Normalized Difference Water Index (NDWI). However, this method relies on a single spectral band and cannot effectively distinguish similar features (e.g., spectral overlap between shadow pixels and water bodies). Furthermore, the threshold requires manual adjustment, making it difficult to adapt to remote sensing data from different regions and time phases. While traditional machine learning methods (such as K-means clustering and support vector machines) can utilize multispectral features, feature extraction relies on manual design (e.g., texture and shape), lacking robustness to the complex spectral variation found in flood scenarios. Furthermore, cluster centers must be pre-set, making them incapable of adapting to data distribution. Furthermore, traditional methods generally lack frequency-domain analysis of spectral features, making it difficult to capture implicit patterns in features, resulting in a "salt-and-pepper" effect in segmentation results and blurred boundary positioning.

[0004] In recent years, deep learning technology has made significant progress in the field of remote sensing semantic segmentation. Typical methods such as U-Net and its improved models automatically learn image features through an encoder-decoder architecture, significantly improving the accuracy of ground object segmentation. Some studies have introduced attention mechanisms to strengthen the weight distribution of spatial and channel features, or to fuse multi-scale features to enhance contextual understanding. However, existing deep learning models still face the following challenges in flood monitoring: First, the spectral information of remote sensing images is not fully utilized. The traditional RGB color space has difficulty separating the independent effects of hue, saturation, and brightness, resulting in insufficient decoupling of feature expression; second, the clustering preprocessing stage often uses a fixed-weight feature fusion strategy, which cannot dynamically balance the contribution of spatial distance and spectral differences to the clustering results; third, the geocoding process relies solely on traditional geometric correction models, without considering the correlation between spectral features and spatial location, resulting in local offsets in the corrected boundaries. In addition, multi-temporal change analysis is usually based on geometric feature matching, ignoring the temporal consistency of spectral features, affecting the reliability of change detection. Summary of the Invention

[0005] Based on the above technical problems, this application discloses a method for rapid remote sensing extraction of flood inundation range based on harmonic spectrum information cutting, including:

[0006] S1. Preprocess the acquired remote sensing image by converting it from RGB color space to HSV color space, extract the spectral information of the image, and decompose the spectral information into harmonic spectral components based on discrete cosine transform;

[0007] S2. Construct an improved clustering model based on the density peak clustering algorithm, input the harmonic spectrum components as clustering features into the model, calculate the local density and distance of the pixels, adaptively determine the cluster center, and cluster the pixels in the remote sensing image into suspected water body and non-water body categories;

[0008] S3. Optimize the clustering results using a semantic segmentation network based on deep learning. The clustered image is used as input. During network training, harmonic spectrum information is introduced as an additional supervisory signal to enhance the network's ability to distinguish between water and non-water features, and the water and non-water segmentation results are output.

[0009] S4. Combine the geographic coordinate information of the remote sensing image to geocode the segmentation results and generate a thematic map containing the flood inundation range. Based on the dynamic time warping algorithm, the thematic maps of different phases are compared and analyzed to quickly obtain the change information of the flood inundation range.

[0010] Preferably, the process of preprocessing the acquired remote sensing image in S1, converting it from RGB color space to HSV color space and extracting spectral information is as follows: for each pixel in the RGB color space, its red, green and blue component values ​​are respectively , through the formula Calculating Hue , among which hour, according to and The size relationship determines the special value, through the formula Calculating saturation ; Through the formula Calculating brightness , convert the image from RGB color space to HSV color space, and extract the value of each channel in the HSV color space as the spectral information of the image.

[0011] Preferably, the extracted spectral information is decomposed into harmonic spectral components in S1, specifically: the channel data of the extracted spectral information in the HSV color space is ,in 、 are the horizontal and vertical coordinates of the image pixels, respectively. Perform a two-dimensional discrete cosine transform to obtain the transform coefficient matrix , the calculation formula is ,in 、 are the width and height of the image, 、 are the corresponding normalization coefficients, and Divided into three sub-coefficient matrices: low frequency, medium frequency, and high frequency , perform inverse two-dimensional discrete cosine transform on the three sub-coefficient matrices respectively to obtain the corresponding harmonic spectrum components , the formula is: ,in .

[0012] Preferably, the improved clustering model based on the density peak clustering algorithm is constructed in S2, specifically: by using the spatial distance weight factor and spectral difference weighting factor , calculate pixel points The local density , the formula is: ,in is the total number of pixels, Pixel and The spatial Euclidean distance of 、 Pixels 、 The eigenvectors of the harmonic spectral components are is the Gaussian kernel parameter; calculate the pixel point distance When, through the formula Determine the weighted minimum distance to the pixel with higher density, and use an adaptive threshold based on the calculated local density and distance Automatically identify cluster centers, where 、 are the corresponding adjustable coefficients, 、 are the means of the local density and distance of all pixels, respectively, and the pixels in the remote sensing image are clustered into suspected water body and non-water body classes.

[0013] Preferably, the spatial distance weight factor and spectral difference weighting factor By constructing a dual objective function based on adaptive weight optimization Perform iterative optimization, where 、 They are the suspected water body and non-water body data sets after clustering, Representation dataset The variance of 、 is the balance coefficient; in each iteration, according to the current and Calculate the dual objective function value and update it through the gradient descent algorithm and , until convergence.

[0014] Preferably, the semantic segmentation network based on deep learning in S3 adopts an improved U-Net architecture with a harmonic spectrum feature fusion module. In the encoding stage, the feature map output by each convolution block is calculated by the formula And the corresponding harmonic spectrum component characteristic diagram Perform weighted fusion, where is the feature map output by the convolution block, 、 are respectively learnable weight parameters; in the decoding stage, after channel splicing the upsampled feature map and the fused feature map of the corresponding layer in the encoding stage, the attention weights of the pixels in the harmonic spectrum feature dimension and the image space dimension are calculated through the attention mechanism module to obtain the weighted feature map; the network loss function adopts weighted cross entropy loss ,in is the total number of pixels, It is obtained by adaptively calculating the entropy value of the harmonic spectrum characteristic of the pixel point. is the true label of the pixel, To predict the probability, output the water body and non-water body segmentation results.

[0015] Preferably, the attention mechanism module calculates the attention weight of the pixel in the harmonic spectrum feature dimension and the image space dimension as follows: the feature map after channel splicing is recorded as ,in is the number of channels, 、 are the height and width of the feature map respectively; Perform global average pooling in the spatial dimension to obtain the harmonic spectrum feature vector , through the formula Calculate the attention weight of the harmonic spectrum feature dimension, where is the learnable weight matrix, is the bias vector, is the Sigmoid activation function; then Grouping is done on the channel dimension, and each group generates a spatial attention map through convolution operation, using the formula Calculate the attention weight of the spatial dimension, where For the Group feature map, is the number of groups, is the convolution operation, To adjust the parameters, the Hadamard product of the harmonic spectrum feature dimension and the attention weight of the image space dimension is performed to obtain the attention weight ,pass Get the weighted feature map.

[0016] Preferably, the process of obtaining the water body and non-water body segmentation results is as follows: inputting the weighted feature map into the last convolution layer of the network, and outputting a feature map with a dimension of 2 , which correspond to the probability distribution of the pixel belonging to water and non-water respectively; using the threshold optimization strategy based on harmonic spectrum similarity, for each pixel , calculate the Euclidean distance between its harmonic spectrum feature vector and the mean vector of the harmonic spectrum feature of water and non-water samples in the training set , through the formula Dynamically generate pixel-level segmentation thresholds, where is the balance parameter; The probability of water body is greater than The pixels of the water body are determined to be water bodies, and the rest are non-water bodies, and the accurate water body and non-water body segmentation results are output.

[0017] Preferably, the process of geocoding the segmentation results in combination with the geographic coordinate information of the remote sensing image in S4 is as follows: constructing a three-dimensional spatial mapping model ,in is the image pixel coordinate, For the corresponding geographic space coordinates, the model is controlled by the set of control points , using the least squares method to solve, where the control points are automatically obtained by SIFT feature matching combined with the ground control point database; the boundary of the water area in the segmentation result is extracted to generate the water body vector boundary ,in is the pixel coordinate of the boundary point; for each boundary point , through the formula Perform geographic coordinate conversion, where is the spatial correction amount based on the harmonic spectrum characteristics, through the formula calculate, For the control points, is the pixel distance, is the radial basis function, is the coordinate residual at the control point, is the weight coefficient; the corrected boundary points Generate a thematic map containing the flood inundation range using the Delaunay triangulation algorithm.

[0018] Preferably, the S4 performs comparative analysis on thematic maps of different time phases based on the dynamic time warping algorithm, specifically: representing the boundaries of flood inundation ranges of different time phases as an ordered point set sequence and , where each point Contains geographic coordinates and harmonic spectrum eigenvectors ; Construct a weighted distance matrix , where the elements , is the geographic space Euclidean distance, is the cosine distance of the harmonic spectrum feature, is the weight coefficient; the optimal path is solved by dynamic programming algorithm , the formula is: , calculate the rate of change of flood inundation range based on the optimal path , the formula is: , so that it can quickly obtain information on changes in flood inundation range.

[0019] Compared with the prior art, the technical solution of this application has the following technical effects:

[0020] This application decomposes spectral information into low-frequency, medium-frequency, and high-frequency harmonic spectral components through HSV color space conversion and discrete cosine transform, breaking through the limitation of traditional spectral analysis that only stays in the spatial domain and enhancing the multi-scale expression capability of ground feature features.

[0021] This application improves the density peak clustering algorithm, introduces a dynamic weight optimization mechanism of spatial distance and spectral difference, and iteratively solves the optimal weight combination through a dual-objective function to solve the problem of fixed feature weights in traditional clustering, thereby improving the initial classification accuracy of suspected water bodies and non-water bodies in complex scenes.

[0022] This application embeds a harmonic spectrum feature fusion module in the U-Net architecture, explicitly models the relationship between spectral frequency domain features and spatial positions through the attention mechanism, and uses harmonic spectrum information as a supervisory signal to optimize the loss function, thereby enhancing the network's ability to distinguish subtle spectral differences.

[0023] In the geocoding process of this application, a spatial correction based on harmonic spectrum characteristics is introduced, and the mapping relationship between spectral characteristics and coordinate residuals is modeled by radial basis functions to improve the geographic positioning accuracy of thematic maps; the dynamic time warping algorithm integrates geographic coordinates and harmonic spectrum feature distances to achieve accurate quantification of multi-temporal flooding range changes.

[0024] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application so that it can be implemented in accordance with the contents of the specification, and to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the following is a detailed description of the preferred embodiment of the present application in conjunction with the accompanying drawings.

[0025] Based on the detailed description of the specific embodiments of the present application in conjunction with the accompanying drawings below, those skilled in the art will become more aware of the above and other objects, advantages and features of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following will briefly introduce the drawings required for the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings without inventive work. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn according to the actual scale.

[0027] Figure 1 This is a flow chart of the method for rapid remote sensing extraction of flood inundation range based on harmonic spectrum information cutting of the present invention;

[0028] Figure 2 Comparison of traditional and improved clustering results in an embodiment of the present invention;

[0029] Figure 3 Comparison of traditional and improved segmentation results in an embodiment of the present invention;

[0030] Figure 4 Schematic diagram of energy distribution of harmonic spectrum components in an embodiment of the present invention;

[0031] Figure 5 is a curve showing the convergence change of the dual objective function in an embodiment of the present invention;

[0032] Figure 6 This is a comparison between traditional and improved clustering tributary extraction in an embodiment of the present invention;

[0033] Figure 7 This is a display of the multi-scale water body segmentation results in an embodiment of the present invention;

[0034] Figure 8 This is a diagram showing the flooded area identification results in an embodiment of the present invention. DETAILED DESCRIPTION

[0035] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. In the following description, specific details such as specific configurations and components are provided only to help fully understand the embodiments of the present application. Therefore, it should be clear to those skilled in the art that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present application. In addition, for clarity and brevity, the description of known functions and structures has been omitted in the embodiments.

[0036] It should be understood that references throughout this specification to "one embodiment" or "this embodiment" mean that a particular feature, structure, or characteristic associated with the embodiment is included in at least one embodiment of the present application. Therefore, the appearance of "one embodiment" or "this embodiment" throughout this specification does not necessarily refer to the same embodiment. Furthermore, these particular features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0037] In addition, the present application may repeat reference numerals and / or letters in different examples. This repetition is for the purpose of simplicity and clarity and does not in itself indicate the relationship between the various embodiments and / or settings discussed.

[0038] The term "and / or" in this article is only a description of the association relationship of associated objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist at the same time. The term " / and" in this article describes another type of association object relationship, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the previous and subsequent associated objects are in an "or" relationship.

[0039] The term "at least one" in this article is merely a description of the association relationship between associated objects, indicating that three relationships may exist. For example, at least one of A and B can mean: A exists alone, A and B exist at the same time, and B exists alone.

[0040] It should also be noted that, in this document, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include," "comprises," or any other variations thereof are intended to cover non-exclusive inclusion.

[0041] Example 1

[0042] This embodiment mainly describes a method for rapid remote sensing extraction of flood inundation range based on harmonic spectrum information cutting. Figure 1 Shown, including:

[0043] S1. Preprocess the acquired remote sensing image by converting it from RGB color space to HSV color space, extract the spectral information of the image, and decompose the spectral information into harmonic spectral components based on discrete cosine transform;

[0044] S2. Construct an improved clustering model based on the density peak clustering algorithm, input the harmonic spectrum components as clustering features into the model, calculate the local density and distance of the pixels, adaptively determine the cluster center, and cluster the pixels in the remote sensing image into suspected water body and non-water body categories;

[0045] S3. Optimize the clustering results using a semantic segmentation network based on deep learning. The clustered image is used as input. During network training, harmonic spectrum information is introduced as an additional supervisory signal to enhance the network's ability to distinguish between water and non-water features, and the water and non-water segmentation results are output.

[0046] S4. Combine the geographic coordinate information of the remote sensing image to geocode the segmentation results and generate a thematic map containing the flood inundation range. Based on the dynamic time warping algorithm, the thematic maps of different phases are compared and analyzed to quickly obtain the change information of the flood inundation range.

[0047] Furthermore, the process of preprocessing the acquired remote sensing image in S1, converting it from RGB color space to HSV color space and extracting spectral information is as follows: for each pixel in RGB color space, its red, green and blue component values ​​are respectively , through the formula Calculating Hue , among which hour, according to and The size relationship determines the special value, through the formula Calculating saturation ; Through the formula Calculating brightness , convert the image from RGB color space to HSV color space, and extract the value of each channel in the HSV color space as the spectral information of the image.

[0048] Furthermore, the extracted spectral information is decomposed into harmonic spectral components in S1, specifically: the channel data of the extracted spectral information in the HSV color space is ,in 、 are the horizontal and vertical coordinates of the image pixels, respectively. Perform a two-dimensional discrete cosine transform to obtain the transform coefficient matrix , the calculation formula is ,in 、 are the width and height of the image, 、 are the corresponding normalization coefficients, and Divided into three sub-coefficient matrices: low frequency, medium frequency, and high frequency , perform inverse two-dimensional discrete cosine transform on the three sub-coefficient matrices respectively to obtain the corresponding harmonic spectrum components , the formula is: ,in .

[0049] Furthermore, the improved clustering model based on the density peak clustering algorithm is constructed in S2, specifically: by using the spatial distance weight factor and spectral difference weighting factor , calculate pixel points The local density , the formula is: ,in is the total number of pixels, Pixel and The spatial Euclidean distance of 、 Pixels 、 The eigenvectors of the harmonic spectral components are is the Gaussian kernel parameter; calculate the pixel point distance When, through the formula Determine the weighted minimum distance to the pixel with higher density, and use an adaptive threshold based on the calculated local density and distance Automatically identify cluster centers, where 、 are the corresponding adjustable coefficients, 、 are the means of the local density and distance of all pixels, respectively, and the pixels in the remote sensing image are clustered into suspected water body and non-water body classes.

[0050] Furthermore, the spatial distance weight factor and spectral difference weighting factor By constructing a dual objective function based on adaptive weight optimization Perform iterative optimization, where 、 They are the suspected water body and non-water body data sets after clustering, Representation dataset The variance of 、 is the balance coefficient; in each iteration, according to the current and Calculate the dual objective function value and update it through the gradient descent algorithm and , until convergence.

[0051] Furthermore, the semantic segmentation network based on deep learning in S3 adopts an improved U-Net architecture with a harmonic spectrum feature fusion module. In the encoding stage, the feature map output by each convolution block is calculated by the formula And the corresponding harmonic spectrum component characteristic diagram Perform weighted fusion, where is the feature map output by the convolution block, 、 are respectively learnable weight parameters; in the decoding stage, after channel splicing the upsampled feature map and the fused feature map of the corresponding layer in the encoding stage, the attention weights of the pixels in the harmonic spectrum feature dimension and the image space dimension are calculated through the attention mechanism module to obtain the weighted feature map; the network loss function adopts weighted cross entropy loss ,in is the total number of pixels, It is obtained by adaptively calculating the entropy value of the harmonic spectrum characteristic of the pixel point. is the true label of the pixel, To predict the probability, output the water body and non-water body segmentation results.

[0052] Furthermore, the attention mechanism module calculates the attention weight of the pixel in the harmonic spectrum feature dimension and the image space dimension as follows: the feature map after channel splicing is recorded as ,in is the number of channels, 、 are the height and width of the feature map respectively; Perform global average pooling in the spatial dimension to obtain the harmonic spectrum feature vector , through the formula Calculate the attention weight of the harmonic spectrum feature dimension, where is the learnable weight matrix, is the bias vector, is the Sigmoid activation function; then Grouping is done on the channel dimension, and each group generates a spatial attention map through convolution operation, using the formula Calculate the attention weight of the spatial dimension, where For the Group feature map, is the number of groups, is the convolution operation, To adjust the parameters, the Hadamard product of the harmonic spectrum feature dimension and the attention weight of the image space dimension is performed to obtain the attention weight ,pass Get the weighted feature map.

[0053] Furthermore, the process of obtaining the water body and non-water body segmentation results is as follows: the weighted feature map is input to the last convolution layer of the network, and the feature map with a dimension of 2 is output. , which correspond to the probability distribution of the pixel belonging to water and non-water respectively; using the threshold optimization strategy based on harmonic spectrum similarity, for each pixel , calculate the Euclidean distance between its harmonic spectrum feature vector and the mean vector of the harmonic spectrum feature of water and non-water samples in the training set , through the formula Dynamically generate pixel-level segmentation thresholds, where is the balance parameter; The probability of water body is greater than The pixels of the water body are determined to be water bodies, and the rest are non-water bodies, and the accurate water body and non-water body segmentation results are output.

[0054] Furthermore, the process of geocoding the segmentation results in combination with the geographic coordinate information of the remote sensing image in S4 is as follows: constructing a three-dimensional spatial mapping model ,in is the image pixel coordinate, For the corresponding geographic space coordinates, the model is controlled by the set of control points , using the least squares method to solve, where the control points are automatically obtained by SIFT feature matching combined with the ground control point database; the boundary of the water area in the segmentation result is extracted to generate the water body vector boundary ,in is the pixel coordinate of the boundary point; for each boundary point , through the formula Perform geographic coordinate conversion, where is the spatial correction amount based on the harmonic spectrum characteristics, through the formula calculate, For the control points, is the pixel distance, is the radial basis function, is the coordinate residual at the control point, is the weight coefficient; the corrected boundary points Generate a thematic map containing the flood inundation range using the Delaunay triangulation algorithm.

[0055] Furthermore, the S4 performs comparative analysis on thematic maps of different time phases based on the dynamic time warping algorithm, specifically: the boundaries of flood inundation ranges of different time phases are represented as ordered point set sequences and , where each point Contains geographic coordinates and harmonic spectrum eigenvectors ; Construct a weighted distance matrix , where the elements , is the geographic space Euclidean distance, is the cosine distance of the harmonic spectrum feature, is the weight coefficient; the optimal path is solved by dynamic programming algorithm , the formula is: , calculate the rate of change of flood inundation range based on the optimal path , the formula is: , so that it can quickly obtain information on changes in flood inundation range.

[0056] This embodiment describes in detail that this application can effectively improve the accuracy and efficiency of flood inundation range extraction, accurately distinguish water bodies from non-water bodies, optimize segmentation results, and geocoding combined with harmonic spectrum correction to improve spatial positioning accuracy through harmonic spectrum information cutting, improved clustering model, deep learning semantic segmentation, and dynamic analysis. It provides high-precision and timely technical support for flood emergency rescue, disaster assessment, etc.

[0057] Based on Example 1, this implementation describes in detail the specific implementation effects of this application, specifically:

[0058] Jiangxi Province, located on the southern bank of the middle and lower reaches of the Yangtze River, boasts a predominantly mountainous and hilly terrain and a well-developed water system. The Poyang Lake basin is a high-risk area for flooding in my country. The study area, Yanxi Town, Taihe County, Ji'an City, Jiangxi Province (115°01′-115°20′E, 26°58′-27°14′N), is located in the middle reaches of the Ganjiang River. This area boasts low-lying terrain and a dense river network. Influenced by a subtropical monsoon climate, summer rainstorms are frequent, and flooding poses a serious threat to local agricultural production and residents' livelihoods.

[0059] This study collected multi-source remote sensing data from June 2023 to August 2024 in the study area, including:

[0060] Sentinel-2L2A multispectral imagery: spatial resolution 10m, spectral range covers visible light to near-infrared bands (B1-B12), a total of 32 valid images were obtained, of which images on July 6, July 12, and July 20, 2023 were used for flood range extraction experiments; Landsat-8OLI / TIRS imagery: spatial resolution 30m, used as auxiliary verification data, 18 images of the same period were obtained; UAV (DJIMavic3) aerial photography data: spatial resolution 0.05m, aerial photography of the core area of ​​the study area (about 50km²) was carried out on July 15, 2023 during the peak period of flooding, to obtain orthophotos (DOM) and digital surface models (DSM).

[0061] Ground Control Points (GCPs): 128 GCPs were evenly distributed across the study area using RTK-GPS (Zhonghaida V90). These points were located at prominent features such as road intersections and field ridge turning points, achieving a planar accuracy of ±2 cm and an elevation accuracy of ±3 cm. Water Sample Collection: Combining drone imagery and field surveys, 2,145 water sample points (including natural water bodies, flooded farmland, and flooded roads) and 3,872 non-water sample points (including bare land, vegetation, and buildings) were annotated in the study area for model training and accuracy verification.

[0062] Meteorological and hydrological data: Data from hydrological stations near the study area in July 2023 provided by the Ji'an Hydrological Bureau were collected, including the real-time water level of the Ganjiang River Yanxi Station (updated at 8:00 and 20:00 daily), the warning water level (82.5m) and the guaranteed water level (84.0m); meteorological data from the same period showed that the cumulative rainfall in the study area in July 2023 reached 486mm, which was 2.46 times the multi-year average monthly rainfall (198mm). The single-day rainfall on July 12 reached 152mm, causing regional flood disasters.

[0063] Data preprocessing is performed on remote sensing images. Radiometric calibration and atmospheric correction are performed: Sen2Corv3.0 is used to perform atmospheric correction on Sentinel-2 images and convert them into surface reflectance data; Landsat-8 images are atmospherically corrected using the FLAASH algorithm; geometric precision correction is performed: based on the 128 GCPs measured by RTK-GPS, a quadratic polynomial transformation (error controlled within 0.5 pixels) is used to perform geometric correction on all remote sensing images, and the unified projection coordinate system is WGS84 / UTMZone50N; HSV color space conversion: RGB images are converted to HSV space, and the three-channel data of hue (H), saturation (S), and value (V) are extracted for subsequent harmonic spectrum analysis.

[0064] UAV data processing was carried out. The UAV aerial photography data was processed using Pix4Dmapper software to generate DOM and DSM with a resolution of 0.05m for the study area. The accuracy verification showed a plane error of ±3cm and an elevation error of ±5cm. Based on the manual interpretation of DOM, a true value map of the flood range was generated as the benchmark data for model verification.

[0065] In order to verify the effectiveness of the technology of this application, the following existing technologies are selected for comparison:

[0066] Traditional spectral threshold method: based on the normalized difference water index (NDWI), the threshold is set at 0.3 (empirical value);

[0067] Traditional density peak clustering (DPClustering): uses Euclidean distance to measure spatial features and does not introduce spectral weight optimization;

[0068] Classic U-Net semantic segmentation model: based on RGB image input, without integrating harmonic spectrum features.

[0069] A spectral feature analysis comparison experiment was conducted. Sentinel-2 images of the study area on July 12, 2023, were selected to compare the ability of RGB space and HSV space to distinguish water bodies from non-water bodies. 1,000 water samples and 1,000 non-water samples were randomly selected, and the inter-class separability (Jeffries-Matusita distance, JM distance) of each channel feature was calculated. The results are shown in Table 1:

[0070] Table 1

[0071] Feature Dimension Water body mean Non-water body mean JM distance RGB-R 0.23±0.05 0.31±0.08 1.23 RGB-G 0.35±0.06 0.42±0.09 1.18 RGB-B 0.28±0.04 0.25±0.07 0.89 HSV-H 0.68±0.12 0.45±0.15 1.89 HSV-S 0.42±0.09 0.28±0.11 1.56 HSV-V 0.55±0.07 0.63±0.06 1.32

[0072] As can be seen from Table 1 above, the H channel (hue) in the HSV space has the strongest inter-class separability (JM distance of 1.89), which is significantly better than the RGB channels, indicating that the hue feature is most critical for distinguishing water bodies from non-water bodies. The JM distance of the RGB-B channel is only 0.89, indicating that the blue light band has difficulty in effectively distinguishing water bodies from shadows (such as building shadows and vegetation shadows). This is consistent with the high misclassification rate of the traditional NDWI method in the study area.

[0073] A clustering accuracy comparison experiment was conducted, comparing traditional density peak clustering with improved density peak clustering. Based on the low-frequency harmonic spectral components of the HSV-V channel (obtained through discrete cosine transform), the study area images were clustered using traditional DPClustering and improved DPClustering (dynamic weight optimization) respectively, and divided into suspected water body and non-water body classes. Evaluation indicators: Rand Index (RI) and Adjusted Rand Index (ARI) were used to assess cluster consistency. The results are shown in Table 2.

[0074] Table 2

[0075] Clustering methods RI ARI Clustering time (min) Traditional DPClustering 0.72±0.04 0.58±0.03 18.2 Improved DPClustering 0.89±0.02 0.81±0.01 22.5

[0076] like Figure 2 As shown in the typical regional comparison, in the farmland area along the Ganjiang River in the study area, traditional DPClustering misclassified a large area of ​​non-flooded farmland as water bodies ( Figure 2 a), the RI value is only 0.65, Figure 1 In a, we can see the clustering results of the traditional DPClustering method in the farmland area along the Ganjiang River. Because the spectral difference weight is not taken into account, a large number of wet farmlands that are not flooded (green areas) are misclassified as water bodies (blue areas). The improved method uses dynamic weight optimization ( =0.52, =0.48) significantly improved the clustering accuracy ( Figure 2 b), RI value reaches 0.91, Figure 2 b shows the clustering results of the improved DPClustering method in the same area. It effectively distinguishes water bodies (blue areas) from moist farmland (green areas) by optimizing the spatial distance and spectral difference weights through a dual-objective function. Analysis shows that the traditional method does not consider the spectral difference weights, resulting in spatially adjacent non-water bodies (such as moist farmland) being mistakenly clustered as water bodies. The improved method balances the contributions of spatial distance and spectral difference through dual-objective function optimization.

[0077] Semantic segmentation accuracy comparison experiment, model training and validation, dataset division: the 6017 labeled samples (2145 water bodies and 3872 non-water bodies) were divided into training set, validation set and test set in a ratio of 7:2:1;

[0078] Training parameters: Both traditional U-Net and Spec2VolCAM U-Net use the Adam optimizer, a learning rate of 1e-4, a batch size of 16, and training for 50 epochs.

[0079] Evaluation indicators include overall accuracy (OA), water body accuracy (PA), non-water body accuracy (NPA), and F1-score. The results are shown in Table 3.

[0080] Table 3

[0081] Clustering methods RI ARI Clustering time (min) Clustering methods Traditional DPClustering 0.72±0.04 0.58±0.03 18.2 Traditional DPClustering Improved DPClustering 0.89±0.02 0.81±0.01 22.5 Improved DPClustering

[0082] Visual comparison, in the core urban area of ​​the study area, such as Figure 3 As shown, the traditional U-Net is good at predicting building shadows ( Figure 3 a red area) and flooded roads, the PA value is only 72%, while Spec2VolCAMU-Net uses the harmonic spectrum feature fusion module ( Figure 3b) Effectively identifying waterlogged areas (blue areas), with the PA value increased to 89%. Further analysis revealed that the high-frequency components of the harmonic spectrum (corresponding to the edges of objects) significantly enhance the network's ability to perceive water boundaries.

[0083] Key improvements and experimental verification of the technology in this application include harmonic spectrum feature extraction and analysis: discrete cosine transform (DCT) parameter optimization, two-dimensional DCT transformation of HSV channel data, and experimental determination of the thresholds for dividing low-frequency, medium-frequency, and high-frequency components:

[0084] like Figure 4 As shown, the low-frequency component: u≤M / 4, v≤N / 4 (M, N are image sizes), accounting for 70% of the total energy, reflecting the overall brightness of the ground object ( Figure 4 a); Intermediate frequency component: M / 4 <u≤3M / 4,N / 4<v≤3N / 4,占总能量的25%,包含地物纹理信息( Figure 4 b); High frequency components: u>3M / 4, v>3N / 4, accounting for 5% of the total energy, corresponding to the edge of the object and noise ( Figure 4 c).

[0085] The harmonic spectrum components express flood characteristics: low-frequency components: appear as uniform low brightness values ​​in flooded areas, contrasting with the high-brightness farmland in non-flooded areas; mid-frequency components: can clearly show the texture differences between water bodies and vegetation, such as the regular texture of rice fields and the smooth texture of water bodies; high-frequency components: highlight the boundary characteristics of water bodies and can be used to refine segmentation results.

[0086] Improved dynamic weight optimization of density peak clustering, dual objective function convergence analysis, dynamic optimization process, objective function F( ) changes with the number of iterations as Figure 5 As shown in the figure, in the first 50 iterations, the intra-class variance (Var(C1)+Var(C2)) decreases rapidly, and stabilizes after the 80th iteration. =0.56±0.02, =0.44±0.01, indicating that the weights of spatial distance and spectral difference reach the optimal balance.

[0087] Comparison of spatial distribution of clustering results. The clustering results of the improved method for the Ganjiang River tributaries in the study area are as follows: Figure 6 As shown, the traditional method ( Figure 6 a) misses the small tributaries with width less than 10m (indicated by arrows), while the improved method ( Figure 6 b) By enhancing the spectral difference weights, the tributary network is completely extracted, and the recall rate is improved from 65% to 92%.

[0088] A deep learning segmentation model with harmonic spectrum fusion was developed to verify the effectiveness of the attention mechanism. It was found that the attention module of Spec2VolCAM U-Net can significantly enhance the feature response of the water area. Quantitative analysis showed that the attention weight entropy of the proposed model is 32% lower than that of the traditional U-Net, indicating that it has stronger feature focusing ability.

[0089] The multi-scale feature fusion effect is achieved by fusing low, medium and high frequency harmonic spectral components in the encoding stage and realizing multi-scale feature fusion through cross-layer connection in the decoding stage. The segmentation results of water bodies of different scales in the study area are shown in Figure 2. Figure 7 As shown, for lakes with an area > 1 km² ( Figure 7 a), the model can accurately extract the boundary; for the water pit with an area of ​​less than 0.1 km² ( Figure 7 b) The recall rate reaches 88%, significantly higher than the 56% of the traditional U-Net.

[0090] The accuracy of geocoding and dynamic analysis is improved. The spatial correction amount based on the harmonic spectrum, the positioning error comparison between the traditional geometric correction (polynomial transformation) and the method of this application are shown in Table 4. In the hilly area of ​​the study area, the average plane error of the traditional method is 4.2m, while after the introduction of the harmonic spectrum correction, the error is reduced to 1.8m, and the elevation error is reduced from 3.5m to 1.2m.

[0091] Table 4

[0092] Calibration method Plane error (m) Elevation error (m) Calibration time (min) Traditional polynomial transformation 4.2±1.5 3.5±1.2 15.6 Harmonic spectrum correction 1.8±0.6 1.2±0.4 22.3

[0093] Multi-phase change detection experiments, such as Figure 8 As shown in the figure, change detection was performed on images taken on July 6, 2023 (before the disaster) and July 12, 2023 (during the disaster). The traditional method misjudged some riverbanks covered with vegetation as flooded areas. However, the method proposed in this application accurately identified the true flooded range through temporal consistency analysis of harmonic spectral features, reducing the missed detection rate from 21% to 7% and the false detection rate from 18% to 5%.

[0094] Comprehensive performance comparison and advantage analysis, as well as accuracy comparison summary, are shown in Table 5:

[0095] Evaluation Metrics Traditional spectral thresholding method Traditional DPClustering Traditional U-Net This application technology Overall accuracy (%) 68.5±2.3 78.2±1.8 82.3±1.2 91.7±0.8 Water extraction recall rate (%) 71.2±3.1 79.5±2.5 83.7±1.9 92.4±1.3 Boundary error (m) 8.5±2.0 6.3±1.5 4.8±1.2 2.1±0.5 Processing time (image / scene) 5min 25min 40min 18min

[0096] In summary of the technical advantages, this application utilizes multi-dimensional feature fusion to overcome the bottleneck of spectral confusion. Through HSV space conversion and harmonic spectrum decomposition, the spectral features are expanded from 3 to 9 dimensions, resolving the spectral confusion between water, shadows, and wet objects in traditional RGB space. In experiments conducted in the study area, the average JM distance between water and non-water bodies increased by 45%, providing a more discriminative feature foundation for subsequent clustering and segmentation.

[0097] This implementation describes in detail the dynamic weight optimization and attention mechanism to enhance the capabilities of joint spatial-spectral modeling. Improved density peak clustering dynamically optimizes spatial and spectral weights through a dual-objective function, increasing the cluster consistency index (ARI) from 0.58 to 0.81, effectively preventing traditional methods from underclassifying small water bodies and misclassifying non-water bodies. The harmonic spectrum feature fusion module and attention mechanism in Spec2VolCAMU-Net improve the network's ability to refine water boundaries by 56%, increasing the F1-score from 0.802 to 0.908, particularly in densely built-up areas and complex terrain. Harmonic spectrum-driven geographic analysis enhances spatiotemporal accuracy. Incorporating harmonic spectrum feature corrections into geocoding reduces positioning error by 57%, meeting the requirements of high-precision emergency mapping (error < 2 meters). A dynamic time warping algorithm integrates spectral and temporal features, reducing the error in multi-temporal change detection from 19.5% to 6.0%, improving processing efficiency by 55%, and enabling rapid quantitative analysis of flood dynamics.

[0098] The above are only preferred embodiments of the present invention, which do not limit the scope of protection of the present invention. For those skilled in the art, the present invention can be modified and varied in various ways. Any changes, modifications, replacements, integrations and parameter changes to these embodiments through conventional substitutions or that can achieve the same functions without departing from the principles and spirit of the present invention fall within the scope of protection of the present invention.

Claims

1. A rapid remote sensing method for extracting flood inundation range based on harmonic spectrum information cutting, characterized by: include: S1. Preprocess the acquired remote sensing image by converting it from RGB color space to HSV color space, extract the spectral information of the image, and decompose the spectral information into harmonic spectral components based on discrete cosine transform; S2. Construct an improved clustering model based on the density peak clustering algorithm, input the harmonic spectrum components as clustering features into the model, calculate the local density and distance of the pixels, adaptively determine the cluster center, and cluster the pixels in the remote sensing image into suspected water body and non-water body categories; S3. Optimize the clustering results using a semantic segmentation network based on deep learning. The clustered image is used as input. During network training, harmonic spectrum information is introduced as an additional supervisory signal to enhance the network's ability to distinguish between water and non-water features, and the water and non-water segmentation results are output. S4. Combine the geographic coordinate information of the remote sensing image to geocode the segmentation results and generate a thematic map containing the flood inundation range. Based on the dynamic time warping algorithm, the thematic maps of different phases are compared and analyzed to quickly obtain the change information of the flood inundation range.

2. The method for rapid remote sensing extraction of flood inundation range based on harmonic spectrum information cutting according to claim 1, characterized in that: The process of preprocessing the acquired remote sensing image in S1, converting it from RGB color space to HSV color space and extracting spectral information is as follows: for each pixel in RGB color space, its red, green and blue component values ​​are respectively , through the formula Calculating Hue , among which hour, according to and The size relationship determines the special value, through the formula Calculating saturation ; Through the formula Calculating brightness , convert the image from RGB color space to HSV color space, and extract the value of each channel in the HSV color space as the spectral information of the image.

3. The method for rapid remote sensing extraction of flood inundation range based on harmonic spectrum information cutting according to claim 1 or 2, characterized in that: In S1, the extracted spectral information is decomposed into harmonic spectral components, specifically: the channel data of the extracted spectral information in the HSV color space is ,in 、 are the horizontal and vertical coordinates of the image pixels, respectively. Perform a two-dimensional discrete cosine transform to obtain the transform coefficient matrix , the calculation formula is ,in 、 are the width and height of the image, 、 are the corresponding normalization coefficients, and Divided into three sub-coefficient matrices: low frequency, medium frequency, and high frequency , perform inverse two-dimensional discrete cosine transform on the three sub-coefficient matrices respectively to obtain the corresponding harmonic spectrum components , the formula is: ,in .

4. The method for rapid remote sensing extraction of flood inundation range based on harmonic spectrum information cutting according to claim 1, characterized in that: In S2, an improved clustering model based on the density peak clustering algorithm is constructed, specifically: by using the spatial distance weight factor and spectral difference weighting factor , calculate pixel points The local density , the formula is: ,in is the total number of pixels, Pixel and The spatial Euclidean distance of 、 Pixels 、 The eigenvectors of the harmonic spectral components are is the Gaussian kernel parameter; calculate the pixel point distance When, through the formula Determine the weighted minimum distance to the pixel with higher density, and use an adaptive threshold based on the calculated local density and distance Automatically identify cluster centers, where 、 are the corresponding adjustable coefficients, 、 are the means of the local density and distance of all pixels, respectively, and the pixels in the remote sensing image are clustered into suspected water body and non-water body classes.

5. The method for rapid remote sensing extraction of flood inundation range based on harmonic spectrum information cutting according to claim 4, characterized in that: The spatial distance weight factor and spectral difference weighting factor By constructing a dual objective function based on adaptive weight optimization Perform iterative optimization, where 、 They are respectively the suspected water body and non-water body data sets after clustering. Representing a dataset The variance of 、 is the balance coefficient; In each iteration, according to the current and Calculate the dual objective function value and update it through the gradient descent algorithm and , until convergence.

6. The method for rapid remote sensing extraction of flood inundation range based on harmonic spectrum information cutting according to claim 1, characterized in that: The semantic segmentation network based on deep learning in S3 adopts an improved U-Net architecture with a harmonic spectrum feature fusion module. In the encoding stage, the feature map output by each convolution block is calculated by the formula And the corresponding harmonic spectrum component characteristic diagram Perform weighted fusion, where is the feature map output by the convolution block, 、 are respectively learnable weight parameters; in the decoding stage, after channel splicing the upsampled feature map and the fused feature map of the corresponding layer in the encoding stage, the attention weights of the pixels in the harmonic spectrum feature dimension and the image space dimension are calculated through the attention mechanism module to obtain the weighted feature map; the network loss function adopts weighted cross entropy loss ,in is the total number of pixels, It is obtained by adaptively calculating the entropy value of the harmonic spectrum characteristic of the pixel point. is the true label of the pixel, To predict the probability, output the water body and non-water body segmentation results.

7. The method for rapid remote sensing extraction of flood inundation range based on harmonic spectrum information cutting according to claim 6, characterized in that: The process of calculating the attention weight of the pixel point in the harmonic spectrum feature dimension and the image space dimension by the attention mechanism module is as follows: the feature map after channel splicing is recorded as ,in is the number of channels, 、 are the height and width of the feature map respectively; Perform global average pooling in the spatial dimension to obtain the harmonic spectrum feature vector , through the formula Calculate the attention weight of the harmonic spectrum feature dimension, where is the learnable weight matrix, is the bias vector, is the Sigmoid activation function; then Grouping is done on the channel dimension, and each group generates a spatial attention map through convolution operation, using the formula Calculate the attention weight of the spatial dimension, where For the Group feature map, is the number of groups, is the convolution operation, To adjust the parameters, the Hadamard product of the harmonic spectrum feature dimension and the attention weight of the image space dimension is performed to obtain the attention weight ,pass Get the weighted feature map.

8. The method for rapid remote sensing extraction of flood inundation range based on harmonic spectrum information cutting according to claim 6, characterized in that: The process of obtaining the water body and non-water body segmentation results is as follows: the weighted feature map is input to the last convolution layer of the network, and the feature map with a dimension of 2 is output. , which correspond to the probability distribution of the pixel belonging to water and non-water respectively; using the threshold optimization strategy based on harmonic spectrum similarity, for each pixel , calculate the Euclidean distance between its harmonic spectrum feature vector and the mean vector of the harmonic spectrum feature of water and non-water samples in the training set , through the formula Dynamically generate pixel-level segmentation thresholds, where is the balance parameter; The probability of water body is greater than The pixels of the water body are determined to be water bodies, and the rest are non-water bodies, and the accurate water body and non-water body segmentation results are output.

9. The method for rapid remote sensing extraction of flood inundation range based on harmonic spectrum information cutting according to claim 1, characterized in that: The process of geocoding the segmentation results in combination with the geographic coordinate information of the remote sensing image in S4 is as follows: constructing a three-dimensional spatial mapping model ,in is the image pixel coordinate, For the corresponding geographic space coordinates, the model is controlled by the set of control points , using the least squares method to solve, where the control points are automatically obtained by SIFT feature matching combined with the ground control point database; the boundary of the water area in the segmentation result is extracted to generate the water body vector boundary ,in is the pixel coordinate of the boundary point; for each boundary point , through the formula Perform geographic coordinate conversion, where is the spatial correction amount based on the harmonic spectrum characteristics, through the formula calculate, For the control points, is the pixel distance, is the radial basis function, is the coordinate residual at the control point, is the weight coefficient; the corrected boundary points Generate a thematic map containing the flood inundation range using the Delaunay triangulation algorithm.

10. The method for rapid remote sensing extraction of flood inundation range based on harmonic spectrum information cutting according to claim 1 or 9, characterized in that: In S4, the thematic maps of different time phases are compared and analyzed based on the dynamic time warping algorithm. Specifically, the boundaries of the flood inundation ranges of different time phases are represented as ordered point set sequences. and , where each point Contains geographic coordinates and harmonic spectrum eigenvectors ; Construct a weighted distance matrix , where the elements , is the geographic space Euclidean distance, is the cosine distance of the harmonic spectrum feature, is the weight coefficient; the optimal path is solved by dynamic programming algorithm , the formula is: , calculate the rate of change of flood inundation range based on the optimal path , the formula is: , so that it can quickly obtain information on changes in flood inundation range.

Images