Remote sensing sewage area identification method and system based on graph structure and multi-stage enhancement

Abstract

The invention relates to the technical field of remote sensing image recognition, in particular to a remote sensing sewage area recognition method and system based on a graph structure and multi-stage enhancement. The method comprises the steps of performing data preprocessing and representation enhancement on an acquired remote sensing image; performing sewage salient region preliminary screening on the enhanced remote sensing image, including abnormal enhancement mapping construction based on local statistical distribution; pollution candidate graph extraction based on spatial structure prior driving; enhancing the response of the stable region based on a structure consistency enhancing mechanism of the polluted region; high-precision segmentation and identification of the sewage area comprises the following steps: constructing a multi-resolution residual pyramid structure; carrying out fine-grained boundary structure modeling and uncertainty suppression; generating a sewage distribution probability graph and optimizing structural consistency; according to the method, the multi-resolution residual pyramid structure is constructed, image context information under different perception scales is fully mined, and the sensitivity and edge integrity of the model to a sewage area under a complex texture background are remarkably enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image recognition, and in particular to a remote sensing sewage area recognition method and system based on graph structure and multi-stage enhancement. Background Art

[0002] With the acceleration of industrialization and the continued expansion of urbanization, water pollution is becoming increasingly serious. This is especially true in typical areas such as urban river networks, industrial clusters, and offshore sewage outlets. Sewage discharge is frequent, complex, and widely distributed, posing a serious threat to ecological and environmental safety and the quality of the human water environment. Remote sensing technology, with its advantages in acquiring water information over a large area, at high frequency, and without contact, has gradually become an important means of sewage identification and monitoring. However, in large-scale remote sensing images, sewage areas often exhibit large scale variations, blurred boundaries, and diverse morphologies, posing a challenge to traditional image analysis methods. More intelligent and automated identification solutions are urgently needed.

[0003] Currently, mainstream wastewater identification methods primarily rely on spectral thresholding, texture feature analysis, traditional image segmentation, or shallow machine learning models for preliminary detection. While these methods can achieve a certain degree of accuracy in certain specific water environments, they have significant limitations. These include sensitivity to complex backgrounds, difficulty adapting to imaging differences between different remote sensing platforms, and a significant drop in recognition performance when interference factors are strong or pollutants are sparse or discontinuously distributed. Furthermore, existing methods often rely on fixed scales and regular windows for processing, lacking the ability to characterize the spatial structure of polluted areas. This results in blurred boundary positioning, poor connectivity, and insufficient interpretability.

[0004] Some studies have attempted to improve the accuracy of remote sensing image segmentation by introducing deep learning models (such as U-Net and DeepLab), using large-scale remote sensing data to train models for more robust feature extraction. However, these methods often focus on the pixel or semantic level, failing to fully exploit the potential spatial dependencies and structural constraints between polluted areas. This makes it difficult to reliably support situations such as irregular boundaries and sparse polluted patches. Furthermore, due to the scale variations and clutter of remote sensing images, conventional networks struggle to fully aggregate multi-scale information, resulting in fragmented and incomplete results.

[0005] On the other hand, while current graph structure modeling methods have shown initial success in other remote sensing applications (such as land cover classification and building detection), they have yet to establish a mature technical path for water pollution identification. Most methods employ fixed graph structures or simple adjacent mapping approaches, lacking cross-level and cross-regional spatial consistency constraints, making it difficult to develop a global collaborative enhancement mechanism. Furthermore, the fuzzy boundaries and drastic morphological variations of polluted areas place higher demands on the robustness of the graph structure, making it difficult for traditional graph neural networks to reliably capture the underlying structural commonalities and differences between regions.

[0006] Therefore, there is an urgent need to propose a remote sensing image sewage identification method that integrates graph structure modeling capabilities, has a spatial collaborative enhancement mechanism, and supports phased expression optimization. It can accurately model the spatial distribution relationship, boundary characteristics and significant structure of polluted areas under complex background interference, and automatically extract high-precision sewage area information from original remote sensing images. It also supports subsequent multi-level output requirements such as spatial mapping, graphic overlay and geographic coordinate export, and helps to upgrade the intelligent environmental monitoring and governance system. Summary of the Invention

[0007] In order to solve the above-mentioned problems existing in the current remote sensing image sewage identification task, such as insufficient recognition accuracy, low efficiency in structural information utilization, weak cross-scale feature expression capability, and lack of a spatial positioning closed-loop mechanism, the present invention provides a remote sensing sewage area identification method and system based on graph structure and multi-stage enhancement.

[0008] In the first aspect, the present invention provides a remote sensing sewage area identification method based on graph structure and multi-stage enhancement, which adopts the following technical solutions: A remote sensing sewage area recognition method based on graph structure and multi-stage enhancement includes: Acquisition of remote sensing images; Perform data preprocessing and representation enhancement on the acquired remote sensing images; The enhanced remote sensing images are used to initially screen for significant areas of polluted water, including the construction of anomaly enhancement maps based on local statistical distribution; extraction of pollution candidate maps driven by spatial structure priors; and enhancement of stable regional responses based on a pollution area structural consistency reinforcement mechanism. High-precision segmentation and identification of sewage areas, including multi-resolution residual pyramid structure construction; fine-grained boundary structure modeling and uncertainty suppression; sewage distribution probability map generation and structural consistency optimization; Spatial relationship modeling of sewage areas based on graph attention, including graph structure construction and node relationship initialization; graph attention network modeling and structure enhancement reasoning; spatial consistency enhancement and segmentation result optimization; Output the results.

[0009] Furthermore, the data preprocessing and characterization enhancement of the acquired remote sensing images include introducing a spectral consistency normalization processing strategy and performing spectral domain normalization operations on the original images to eliminate spectral drift caused by non-target factors. In order to ensure comparability between bands at the same scale, the spectral vectors of all pixels are first normalized, and the Z-score normalization form is adopted to convert the pixel values ​​of each band into a form with a mean of 0 and a standard deviation of 1. At the same time, in order to enhance the difference expression and feature differentiation ability between the bands, a learnable frequency domain transformation enhancement mechanism is introduced, and the spectral vector is regarded as a one-dimensional signal, and a Fourier transform is performed on it to extract frequency domain features. In order to suppress frequency domain interference and enhance texture expression, a frequency domain enhancement mechanism based on Fourier transform is introduced to achieve the retention of high-frequency structure and the suppression of low-frequency redundancy.

[0010] Furthermore, the abnormal enhancement mapping based on local statistical distribution is constructed by converting the preprocessed and enhanced remote sensing image tensor ,in 、 are the height and width of the image, For the number of bands, an abnormal enhancement mapping method based on the deviation of local statistical distribution is constructed; for each pixel position in the image , extract it in is the center and the side length is Local window , calculate the mean vector of all pixels in the window in the spectral dimension and the covariance matrix : , , in Indicates the position in the local window The spectral vector of Indicates the number of pixels in the local window.

[0011] Furthermore, the extraction of the pollution candidate map based on spatial structure prior drive includes preprocessing the spectral enhancement image tensor As input, by the prior abnormal response graph Perform weight sorting to obtain , then In the subjectively important band The structure is extracted on the image; the Sobel operator is used to extract the local gradient response of the image in the horizontal and vertical directions, and the overall gradient amplitude map is constructed as the local intensity change response. In order to identify the fuzzy boundary area, the local Laplace transform is introduced to characterize the edge sharpness, and the edge fuzziness scoring function is defined. To highlight the edge abnormality of the contaminated area, the gradient amplitude map is finally and fuzziness score graph Perform weighted fusion to construct a structural saliency map , by setting the significance threshold Binarize the fusion image to extract the pollution candidate area and structural saliency map Expressed as: , The fusion coefficient This can be obtained by adjusting the parameters of the validation set, and the recommended initial value is 0.6.

[0012] Furthermore, the structure consistency enhancement mechanism based on the pollution region enhances the stable region response, including the structure candidate map As the initial saliency hint, based on the preprocessed image tensor A structural feature representation is constructed for cross-temporal alignment and consistency scoring. To measure the structural consistency between the current frame and the reference frame within the candidate region, the vector angle cosine similarity is introduced as a consistency metric. Subsequently, to avoid noise propagation caused by directly using the original structure map, only the structural consistency within the candidate region is statistically analyzed to define the final structural consistency enhancement mask.

[0013] Furthermore, the fine-grained boundary structure modeling and uncertainty suppression include fusion feature maps output by multi-resolution residual pyramids. As input, combined with the original spectral image , in the input image Apply Sobel filter to extract spatial gradient intensity and obtain boundary response map , then the boundary response map Intermediate scale features with pyramids Concatenate in the channel dimension and input a lightweight convolutional attention module to extract boundary saliency weights: , in Indicates channel splicing, is a 1×1 convolution operation, is the Sigmoid activation function.

[0014] Furthermore, the sewage distribution probability map generation and structural consistency optimization include, on the basis of completing multi-scale structure fusion and boundary refinement modeling, introducing the pollution probability map generation and consistency optimization mechanism based on the structural prior map, outputting the final prediction result with continuous space, smooth boundaries and reasonable structure, among which, the fusion feature map output based on the fine-grained boundary structure modeling is , through a convolutional prediction head Generate the probability value P of each pixel belonging to the sewage area, and introduce the consistency regularization term in the training stage The predicted probability map is structurally guided to match the candidate region structure, which is expressed as: , in represents the consistency regularization term, Indicates the height of the image, Indicates the width of the image, Represents the final structural consistency enhancement map of all pixels, Indicates the probability value of all pixels belonging to the sewage area, Represents a very small positive number used to prevent the denominator from being 0.

[0015] Furthermore, the graph attention network modeling and structure enhancement reasoning include introducing a graph attention network, dynamically adjusting the information transmission intensity between adjacent nodes through a multi-head self-attention mechanism, thereby achieving cross-region feature enhancement reasoning, wherein the constructed node set is , the initial feature of each node is , the adjacency weight matrix is , No. Attention head to node The update is expressed as: , in Indicates the Nodes in the head For Node The attention weight, is the characteristic linear transformation matrix, is the attention parameter, represents vector concatenation, For nodes LeakyReLU is a nonlinear activation function used to enhance the model's ability to express negative weights.

[0016] Furthermore, the spatial consistency enhancement and segmentation result optimization include the introduction of graph enhancement mapping mechanism and structure preservation regularization term to improve the spatial consistency and expression integrity of the final segmentation graph, wherein the node features output by the graph attention network are , which means the first The semantic representation of the polluted sub-blocks is then used to assign each node feature to its corresponding pixel set through the region inverse mapping operation. , and get the enhanced feature map ; Then enhance the feature map and fusion feature Perform splicing and fusion and input the lightweight decoder layer To generate the final pollution area prediction probability map In order to alleviate the potential risk of feature offset in the graph reasoning process, a structure-preserving regularization term is introduced to measure the final predicted graph Attention input probability map with the original map The spatial structural consistency between them is defined as the weighted KL divergence regularization term: , in, is the edge saliency map, defined in the previous module, which is used to enforce the consistency constraints on the boundary area.

[0017] The second aspect is a remote sensing sewage area identification system based on graph structure and multi-stage enhancement, including: The data acquisition module is configured to acquire remote sensing images; The preprocessing module is configured to perform data preprocessing and representation enhancement on the acquired remote sensing image; The initial screening module is configured to perform initial screening of polluted areas in enhanced remote sensing images, including the construction of anomaly enhancement maps based on local statistical distribution; extraction of pollution candidate maps driven by spatial structure priors; and enhancement of stable area responses based on a pollution area structural consistency reinforcement mechanism. The recognition module is configured for high-precision segmentation and recognition of sewage areas, including multi-resolution residual pyramid structure construction; fine-grained boundary structure modeling and uncertainty suppression; sewage distribution probability map generation and structural consistency optimization; The modeling module is configured to model the spatial relationship of sewage areas based on graph attention, including graph structure construction and node relationship initialization; graph attention network modeling and structure enhancement reasoning; spatial consistency enhancement and segmentation result optimization; The output module is configured to output the results.

[0018] In a third aspect, the present invention provides a computer-readable storage medium storing a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, a remote sensing sewage area identification method based on graph structure and multi-stage enhancement.

[0019] In a fourth aspect, the present invention provides a terminal device comprising a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, and the instructions are suitable for being loaded and executed by the processor to implement the remote sensing sewage area identification method based on graph structure and multi-stage enhancement.

[0020] In summary, the present invention has the following beneficial technical effects: Compared with the existing remote sensing image sewage area recognition methods that have weak multi-scale feature extraction capabilities, poor spatial consistency, unstable boundary recognition and lack of geographic registration support, the remote sensing image sewage area recognition method proposed in this invention based on graph structure modeling and multi-stage enhancement significantly improves the recognition accuracy, boundary expression ability and geographic projection accuracy of polluted areas in complex remote sensing scenes.

[0021] First, the present invention constructs a multi-resolution residual pyramid structure to fully exploit the image context information at different perception scales, significantly enhancing the model's sensitivity to sewage areas under complex texture backgrounds and edge integrity; then, combining fine-grained edge saliency map modeling with uncertainty suppression mechanism, the edge saliency response weight is introduced when determining boundary areas, and through multi-scale disturbance perception and confidence constraints, background interference and misjudgment of fuzzy areas are effectively suppressed; further, through graph structure modeling and upstream and downstream information constraint mechanism, the spatial topological relationship of polluted water bodies is modeled and corrected, effectively enhancing the consistent expression of regional structure; finally, the output module is designed to accurately map the recognition result mask map to the image and geographic space coordinate system, providing data support for actual pollution control and regulatory applications.

[0022] In a test set of typical remote sensing pollution images, the proposed method improved the average identification accuracy (IoU) of polluted areas from 81.3% compared to traditional methods to 92.7%. The boundary structure integrity score increased from 72.6% to 87.9%, significantly improving the representation of regional edges. The false alarm rate in uncertain areas was reduced from 15.4% to 5.8%, and the spatial projection error of the identification results was reduced from 6.7 meters to 3.2 meters. The system's overall average inference time was 0.74 seconds per image, demonstrating excellent timeliness and deployment stability. With its advantages of high recognition accuracy, strong boundary stability, accurate geographic mapping, and excellent system response efficiency, this method achieves highly reliable identification of polluted areas in remote sensing scenarios, possessing significant engineering practical value and broad potential for widespread adoption. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a schematic diagram of a remote sensing sewage area identification method based on graph structure and multi-stage enhancement according to Example 1 of the present invention.

[0024] Figure 2 This is a graph showing the comparison results of various methods in Example 1 of the present invention in terms of IoU and BCR indicators.

[0025] Figure 3 This is a graph showing the comparative results of various methods in Example 1 of the present invention in terms of FAR-U, SPE, and Time indicators. DETAILED DESCRIPTION

[0026] The present invention will be further described in detail below with reference to the accompanying drawings.

[0027] Example 1 Reference Figure 1 In this embodiment, a remote sensing sewage area identification method based on graph structure and multi-stage enhancement includes: (1) Data preprocessing and representation enhancement module, 1) Spectral consistency normalization processing, In the actual acquisition of remote sensing images, due to external factors such as imaging sensor performance, imaging angle differences, atmospheric conditions, and surface reflectivity, remote sensing images of the same area acquired at different times or using different equipment often exhibit inconsistent spectral responses. This spectral variation can lead to significant differences in the spectral appearance of the same ground feature in different images, thus interfering with the subsequent identification, modeling, and analysis of wastewater areas, and reducing the model's stability and generalization capabilities.

[0028] To solve the above problems, this module introduces a spectral consistency normalization processing strategy to perform spectral domain normalization on the original image to eliminate spectral drift caused by non-target factors and improve the contrast consistency and stability between different bands.

[0029] Assume that the original remote sensing image is represented as a three-dimensional tensor: , in, 、 Represent the height and width of the image respectively, Indicates the number of bands. pixels, and the corresponding multispectral vector can be expressed as: , To ensure comparability between bands at the same scale, the spectral vectors of all pixels are first normalized. Using the Z-score normalization form, the pixel values ​​of each band are converted to a form with a mean of 0 and a standard deviation of 1: , in, and Respectively represent The mean and standard deviation of each band in the entire image, Represents the normalized pixel value.

[0030] At the same time, in order to further enhance the differential expression and feature differentiation capabilities between bands, a learnable frequency domain transformation enhancement mechanism is introduced. The spectral vector is regarded as a one-dimensional signal and Fourier transform is performed on it to extract frequency domain features: , in represents the Fourier transform operation, In order to suppress high-frequency noise while retaining the main spectral structure information, a frequency domain weight mask is designed. , used to enhance or weaken the corresponding frequency band: , Finally, the enhanced time domain representation is restored by inverse Fourier transform: , in represents the enhanced time domain representation, represents the inverse Fourier transform operation, Enhance or weaken the corresponding frequency band. represents the Fourier transform operation, Represents all pixel values ​​after normalization, Represents the frequency domain weight mask.

[0031] After the above processing, the spectral vectors of all pixels are mapped to a feature space with a more stable spectral structure and more robust to abnormal bands, thereby providing a more reliable input representation for subsequent region recognition and modeling.

[0032] 2) Frequency domain shadow interference suppression and texture enhancement, Remote sensing images are often affected by factors such as shadows and strong reflections during the natural imaging process. This is especially true in densely built-up areas such as cities and industrial zones. The shadows cast by large buildings or three-dimensional structures at different solar altitudes can significantly interfere with the image's spectral expression and spatial texture distribution, leading to misidentification of polluted areas or missed detections. Furthermore, local texture distortion in remote sensing images, caused by differences in sensor response and ground material, can also obscure the discriminative characteristics of sewage areas at the spatial structural level.

[0033] Therefore, in order to further suppress frequency domain interference and enhance texture expression, this module introduces a frequency domain enhancement mechanism based on Fourier transform on the basis of spectral consistency normalization processing to achieve the retention of high-frequency structure and the suppression of low-frequency redundancy.

[0034] Specifically, the normalized image obtained after the spectral consistency normalization process in the previous step is: , in, 、 、 are the height, width and number of spectral channels of the image respectively. We first calculate the Perform a two-dimensional Fourier transform to obtain its spectral representation: , in Indicates the Perform a two-dimensional Fourier transform operation on each pixel. Indicates a two-dimensional Fourier transform operation for each spectral channel. Indicates the height of the image, Indicates the width of the image.

[0035] After transforming all channels, the overall spectrum expression can be obtained: , in Indicates the height of the image, Indicates the width of the image, represents the number of spectral channels, Represents the result of Fourier transform operation on each channel. Represents the overall spectrum representation after all channels are transformed.

[0036] In the spectrum, the low-frequency components in the image mainly correspond to the overall brightness and large-scale regional gradients, which are often greatly affected by shadow interference; while the high-frequency components carry structural information such as image edges and textures, which are important for identifying sewage areas. Therefore, constructing a frequency domain mask , to keep only the high-frequency regions:

[0037] in is the frequency domain high-pass filter radius. The suppression result after applying the frequency domain mask is: , in represents the frequency domain mask, represents the image after spectral consistency normalization, represents the Fourier transform operation, Represents the suppression result after applying the frequency domain mask.

[0038] Perform inverse Fourier transform on the frequency domain high-frequency image to obtain the enhanced image expression: , in Indicates the height of the image, Indicates the width of the image, represents the number of spectral channels, represents the inverse Fourier transform operation, Represents the enhanced image expression.

[0039] In order to further improve the distribution characteristics of texture perception, we introduce the texture enhancement factor , weighted fusion is performed on the original image and the frequency domain enhanced image, and the final enhanced image is expressed as: , in The original input comes from the "spectral consistency normalization process" to ensure the consistency of input and output between modules; It can be automatically set based on empirical values ​​or statistical data. This fusion strategy not only preserves the overall spectral information of the image, but also significantly enhances the edge and texture features of the contaminated area, facilitating the subsequent accurate segmentation of the salient area.

[0040] 3) Dynamic band selection and information redundancy suppression, Remote sensing images, especially hyperspectral images, contain a large number of bands (up to hundreds). While this provides rich spectral information for precise identification, it also presents serious problems of "curse of dimensionality" and "redundant interference." On the one hand, many bands are highly correlated, leading to redundant information duplication and inefficient feature expression. On the other hand, specific bands may be affected by atmospheric scattering, imaging noise, or physical obstructions (such as clouds and shadows), becoming "contaminated bands." This introduces interference into the model and affects the accurate identification of wastewater areas.

[0041] To this end, this module proposes a dynamic band selection mechanism based on information contribution to maximize effective information and suppress redundant bands, improving overall representation capabilities and model inference efficiency. This mechanism not only considers the information content of each band in the current image but also incorporates the spectral heterogeneity of features after frequency domain enhancement, resulting in a band selection method that is more adaptable to the task objectives.

[0042] The image tensor after Fourier texture enhancement in the previous stage is: , in, 、 represents the image space dimension, Indicates the number of bands, Indicates the Two-dimensional feature map of the band.

[0043] Then define the information entropy of each band As an indicator of its independent information contribution, the formula is as follows: , in, Indicates band Previous The probability density corresponding to the pixel value can be obtained by histogram normalization estimation. Indicates the total number of intervals in which pixel values ​​are discretized. The higher the information entropy, the richer the discriminant information contained in the band, and the more valuable it is to select.

[0044] In order to further measure the mutual information redundancy between bands, the mutual information matrix is ​​introduced : , in Indicates band With band Combination of pixel values The joint probability of and represent marginal probabilities respectively.

[0045] Next, we introduce the spectral selection score function As the final band evaluation criteria: , This score combines the information contribution of the band itself and the redundancy of other bands. The larger the value, the more worthy the band is to be retained. of To sort, select The bands constitute the optimized band set: , The final optimized image tensor for subsequent processing is: , It is worth noting that the Fourier enhancement With higher frequency texture and edge definition, it can be used in the above process It is more discriminative, avoids the misselection of low-frequency redundant bands, and improves the pertinence and robustness of band selection.

[0046] Through this dynamic band selection mechanism, not only can the data dimension be effectively compressed and the model training and inference efficiency be improved, but also the noise bands can be eliminated and the effective bands can be enhanced, thus laying a high-quality data foundation for the subsequent spatial significance analysis and fine segmentation of sewage areas.

[0047] (2) Sewage significant area initial screening module, 1) Construction of anomaly enhancement maps based on local statistical distribution, In remote sensing image analysis, water bodies and sewage-contaminated areas often exhibit significant differences in spectral and textural characteristics. Particularly at the local scale, polluted areas, affected by chemical emissions or structural disturbances, often exhibit anomalies such as spectral shifts, fragmented textures, and abnormally high or low reflectance. These anomalies are easily diluted by traditional global statistical features. Therefore, it is necessary to introduce a local modeling method with spatial awareness to enhance the potential contaminated areas a priori, thereby improving the sensitivity of the subsequent segmentation module to detect significantly contaminated areas.

[0048] To this end, this module is based on the remote sensing image tensor that has been preprocessed and enhanced in the previous stage. ,in 、 are the height and width of the image, For the number of bands, an anomaly enhancement mapping method based on the deviation of local statistical distribution is constructed.

[0049] For each pixel position in the image , extract it in is the center and the side length is Local window , calculate the mean vector of all pixels in the window in the spectral dimension and the covariance matrix : , , in Indicates the position in the local window The spectral vector of Indicates the number of pixels in the local window.

[0050] Then, the Mahalanobis distance is used to measure the deviation between the current pixel and its local statistical distribution: , in, Represents pixel points The degree of abnormality relative to the statistical characteristics of its neighborhood. The larger the value, the less consistent the point is with the statistical distribution of its local environment and the more likely it is to belong to a potential pollution area.

[0051] Finally, the Mahalanobis distance values ​​of all pixels in the entire image are used to form a pollution anomaly response map , and normalize it: , Normalized anomaly map It can serve as an important priori prompt channel for the subsequent sewage segmentation module, enhance the network's perception of potential polluted areas, and provide key support for refined segmentation and boundary modeling.

[0052] 2) Extraction of contaminated candidate graphs driven by spatial structure priors, Polluted areas in remote sensing images not only exhibit anomalous reflectance characteristics in spectral space but also exhibit significant differences in spatial structure. Compared to clean water, which typically exhibits clear boundaries and smooth textures, polluted waters tend to exhibit local structural inconsistencies, such as blurred boundaries, irregular shapes, and dramatic internal intensity fluctuations. These spatial structural differences reflect the disturbances to the continuity and surface state of the water body caused by the pollution process and are an important basis for constructing spatial priors.

[0053] In order to further extract the candidate pollution regions with significant structure, this module designs a spatial significance construction method that integrates the image gradient response and edge fuzziness score to reveal the structural abnormality of the pollution region at the spatial distribution level. First, the spectral enhancement image tensor output by the preprocessing module is used to As input, by the prior abnormal response graph Perform weight sorting to obtain , then In the subjectively important band Structural extraction is performed on . Let the corresponding two-dimensional image be .

[0054] The classic Sobel operator is used to extract the local gradient response of the image in the horizontal and vertical directions, which is defined as: , Then construct the overall gradient magnitude map as a response to local intensity changes: , The larger the gradient amplitude, the more significant the texture change or edge mutation in the local area, which may correspond to the pollution disturbance area.

[0055] In order to further identify the fuzzy boundary area, the local Laplace transform is introduced to characterize the edge sharpness. The response at is: , in Denotes the discrete Laplace operator. Define the edge fuzziness score function for: , This rating is Normalizes the range, with larger values ​​indicating more blurred boundaries and greater uncertainty. This effectively suppresses the interference of smooth edges in the background water and highlights the edge anomalies of the polluted area.

[0056] In order to construct the spatial structure saliency fusion map, the gradient amplitude map and fuzziness score graph Perform weighted fusion to construct a structural saliency map : , The fusion coefficient It can be obtained by adjusting the parameters of the validation set, and the recommended initial value is 0.6. Then a binary mask is used to extract the polluted candidate area. By setting the significance threshold Perform a binarization operation on the fusion image to extract the pollution candidate area: , Finally, we get the candidate mask , as the regional attention constraint for subsequent sewage fine segmentation and upstream and downstream modeling modules, it reduces the interference of redundant background and improves the processing efficiency and accuracy.

[0057] 3) Strengthening mechanism of structural consistency of polluted areas, In multi-temporal observations of remote sensing images, the spectral and spatial structural characteristics of polluted areas can vary significantly over time, depending on the camera angle, weather conditions, and other factors. This can cause the same polluted area to appear shifted in structure, with blurred boundaries or inconsistent responses at different time points. Without modeling and correction, this can easily lead to false positives and false negatives. In particular, isolated outliers or background disturbances can disrupt the consistent representation of regional boundaries and reduce model stability.

[0058] To improve the structural stability and continuity of the preliminary candidate regions in the temporal dimension, this module proposes a temporal consistency enhancement mechanism based on the similarity of structural features. Through structural alignment operations between multi-temporal images, it enhances the response of stable regions and suppresses the influence of structural drift regions.

[0059] This module outputs the structure candidate graph of the previous module. As the initial saliency hint, based on the preprocessed image tensor Construct structural feature representation for cross-temporal alignment and consistency scoring. Suppose the current phase With reference phase The structural candidate graphs are: , Then the input image tensor is extracted through the structural feature extraction function Get a structural representation: , in The dimension representing the structural features, such as the Sobel edge map, Laplacian map, or texture direction features.

[0060] In order to measure the structural consistency between the current frame and the reference frame in the candidate region, the vector angle cosine similarity is introduced as a consistency metric: , in represents the pixel position in the image, represents the vector inner product, To prevent the small constant from dividing by zero. Then, in order to avoid the noise propagation caused by directly using the original structure map, only the structural consistency within the candidate area is counted, and the final structural consistency enhancement mask is defined: , If the structural similarity of a region is low (for example, due to texture mutation caused by specular reflection, sensor noise, or cloud occlusion), its saliency response will be attenuated. On the contrary, if the regional structure remains consistent in two frames, its response will be maintained or enhanced. The final structural consistency enhancement image obtained after the above processing is: , It will serve as input prompts for the subsequent fine segmentation and pollution attribute modeling modules to guide the fine segmentation processing of structural continuity areas.

[0061] (3) Sewage area high-precision segmentation and recognition module, 1) Multi-resolution residual pyramid structure construction, In remote sensing images, wastewater areas often exhibit complex morphological features such as multi-scale, low-contrast, and blurred boundaries. Traditional single-scale feature extraction methods struggle to simultaneously capture fine-grained edges and large-scale background differences, limiting high-precision segmentation performance. To address this, this module introduces a multi-resolution residual pyramid (MRP) architecture to enhance the network's ability to represent wastewater areas at different spatial scales and achieve a fusion of structural details and contextual semantic information.

[0062] The input of this module is the structural consistency enhancement obtained by the initial screening module. Figure 2 dimensional tensor ,in and are the height and width of the image respectively. We first Constructing multi-layer pyramid feature sequences , among which Layer Pyramid By downsampling factor Generated from the original image, that is: , In order to enhance the nonlinear expression capability of each scale, the residual encoding block (REB) is introduced to perform residual enhancement on each pyramid feature. The layer input features are , then its output after the residual module is: , in is a learnable convolution kernel, represents the convolution operation, is the ReLU activation function, The residual connection is used to preserve the local details of the original input. Then, the residual feature maps of different scales are fused from bottom to top to construct a unified high-dimensional multi-scale structure representation. : , All the smaller-scale feature maps All channels are restored to their original resolution through upsampling and then spliced. The final output fusion feature map It has rich hierarchical structure information and can provide feature support with strong discriminability and spatial consistency for subsequent processing.

[0063] 2) Fine-grained boundary structure modeling and uncertainty suppression, In remote sensing scenes, the edges of sewage areas often exhibit high grayscale transitions and local structural fuzzy features. Especially when affected by interference such as light, wind and waves, the boundary between the polluted area and the background water body is blurred, causing traditional segmentation methods to easily produce errors such as "false contours" or "fragmentation" in the edge areas.

[0064] To improve the model's ability to identify structures in boundary areas, this module proposes a fine-grained structure modeling mechanism based on gradient perception and boundary attention guidance, which dynamically suppresses the prediction uncertainty of edge transition areas while preserving spatial continuity.

[0065] This module is the fusion feature map output by the previous stage multi-resolution residual pyramid module As input, combined with the original spectral image , perform the following operations: First, in the input image Apply Sobel filter to extract spatial gradient intensity and obtain boundary response map : , in, , is the horizontal and vertical convolution kernel of the Sobel operator, Represents a two-dimensional convolution operation. Then the boundary response map Intermediate scale features with pyramids Concatenate in the channel dimension and input a lightweight convolutional attention module to extract boundary saliency weights: , in Indicates channel splicing, is a 1×1 convolution operation, is the Sigmoid activation function. Characterize the boundary structure saliency weight of each pixel. Then use the boundary attention map to perform weighted adjustment on the original fusion feature map, thereby strengthening the discriminant features of the boundary area and suppressing the response disturbance of the structural uncertainty area: , in This weighted strategy essentially implements the mechanism of enhancing boundary regions (attention boost) and suppressing non-boundary regions (residual skip), making the model more accurate in identifying the edges of contaminated regions.

[0066] 3) Generation of sewage distribution probability map and structural consistency optimization, The spatial distribution of wastewater areas in remote sensing images often exhibits structural discontinuities, blurred boundaries, and localized mutations. Direct pixel-by-pixel discrimination based on feature maps can easily lead to the expansion of artifact areas or fragmentation of target areas. To address this, this module, building on multi-scale structural fusion and boundary refinement modeling, introduces a pollution probability map generation and consistency optimization mechanism based on a structural prior map, outputting a final prediction result that is spatially continuous, has smooth boundaries, and a reasonable structure.

[0067] Fusion feature map based on fine-grained boundary structure modeling output , will pass a lightweight convolutional prediction head Generate the probability value of each pixel belonging to the sewage area: , in Represents pixels The confidence level of belonging to the contaminated area, It can be composed of a set of continuous 1×1 convolutions and Sigmoid activations. Considering that a structurally stable candidate mask has been generated in the initial screening stage, , we introduce a consistency regularization term during the training phase Structural guidance is performed on the predicted probability map to match the candidate region structure: , Furthermore, in order to alleviate the ambiguity and instability of boundary area prediction, the boundary attention weight map generated in the previous module is fused , dynamic confidence adjustment of the preliminary prediction results: , in Represents a mild spatial Gaussian smoothing operator, which is used to suppress local noise and improve edge consistency. The resulting structure optimization pollution probability map It will serve as the input of the next stage's "Spatial Relationship Modeling of Wastewater Areas Based on Graph Attention" module, providing a high-confidence semantic basis for spatial mapping of polluted areas and upstream and downstream interactions.

[0068] (4) Sewage area spatial relationship modeling module based on graph attention, 1) Graph structure construction and node relationship initialization, In remote sensing images, sewage areas often exhibit strong spatial coherence, distinct localized clusters, and complex structural morphology. This distribution is influenced by non-Euclidean spatial factors such as terrain slope, water flow direction, and human drainage pathways, and is not random and independent. Therefore, traditional independent prediction methods based on pixels or local windows struggle to capture the deep structural dependencies between regions, making them prone to misjudgments such as prediction gaps, boundary jumps, and void expansion.

[0069] To enhance the model's ability to model cross-regional spatial structures, this module proposes a spatial relationship modeling method based on graph structure modeling and attention mechanisms. It divides the polluted regions into structured nodes, and uses graph structures to represent their spatial and semantic connections, thereby capturing non-local inter-regional dependencies.

[0070] The pollution probability map output by the previous module As the input part, first set the probability threshold Get the pseudo pollution mask map : , Subsequently, based on Perform the connected area extraction operation, define each connected pollution block as a graph node, and construct a node set . Each node Corresponding to a pseudo-contaminated sub-region, its initial feature vector is fused with the feature map in the region The pixel mean is calculated as: , in, Representation node The corresponding pixel coordinate set. Secondly, the edge weight of the graph is determined by the spatial position and semantic feature similarity between nodes, defining the adjacency weight matrix of the graph : , in, Representation node The coordinates of the center of mass in space, As its initial feature representation, and These are adjustment parameters for spatial distance and semantic distance, respectively, used to control the influence of different factors on edge weight calculation. This node initialization and adjacency construction method provides a structural prior and feature foundation for subsequent graph attention network modeling and structure-enhanced reasoning.

[0071] 2) Graph Attention Network Modeling and Structure Enhanced Reasoning, While the aforementioned graph structure has effectively established spatial and semantic connections between wastewater regions, inter-node dependencies still need to be further modeled through explicit information exchange mechanisms. Traditional graph convolutional methods rely on static adjacency structures for weight distribution, making it difficult to fully exploit semantic differences and boundary uncertainty across regions. In remote sensing imagery, different wastewater regions may exhibit distant but semantically similar structural features due to factors such as terrain occlusion and illumination variations. Therefore, it is urgent to introduce a mechanism with adaptive mapping capabilities to enhance the representation of structural consistency.

[0072] To this end, this module introduces the Graph Attention Network (GAT), which dynamically adjusts the information transmission intensity between adjacent nodes through a multi-head self-attention mechanism, thereby achieving cross-region feature enhancement reasoning.

[0073] The node set constructed previously is , the initial feature of each node is , the adjacency weight matrix is , No. Attention head to node The update is expressed as: , in Indicates the Nodes in the head For Node The attention weight, is the characteristic linear transformation matrix, is the attention parameter, represents vector concatenation, For nodes The set of adjacent nodes, LeakyReLU is a nonlinear activation function used to enhance the model’s ability to express negative weights. The outputs of the attention heads are concatenated to obtain the updated node representation: , The above attention mechanism allows the model to automatically assign information propagation weights based on the relative similarity of node features, thereby establishing strong connections between distant but semantically related regions in the graph space, alleviating problems such as disconnection and missed detection, and enhancing the ability to model structural consistency across regions. The final node features It will serve as a structurally enhanced representation, providing a structure-aware semantic basis for subsequent full-image integration and mask map remapping.

[0074] 3) Enhanced spatial consistency and optimized segmentation results, Although the graph attention mechanism can capture the latent semantic relationships between distant regions at the structural level, its operational units are still primarily region-level nodes, making it difficult to directly enforce consistency constraints on the fine-grained spatial pixel distribution of the original image. Therefore, to effectively integrate the graph structure modeling results with the spatial representation of the original image, this module introduces a graph enhancement mapping mechanism and a structure-preserving regularization term to improve the spatial consistency and expression integrity of the final segmentation map.

[0075] The node features output by the graph attention network are , which means the first The semantic representation of the polluted sub-blocks is then used to assign each node feature to its corresponding pixel set through the region inverse mapping operation. , and get the enhanced feature map : , Then the enhanced feature map is combined with the fusion feature Perform splicing and fusion and input the lightweight decoder layer To generate the final pollution area prediction probability map : , in In order to alleviate the potential risk of feature offset during graph reasoning, a structure-preserving regularization term is further introduced to measure the final predicted graph. Attention input probability map with the original map The spatial structural consistency between them. The structural consistency regularization term is defined as the weighted KL divergence form: , in, is the edge saliency map, defined in the previous module, used to strengthen the constraint on the consistency of the boundary area. The final output probability map At the same time, it integrates the collaborative reasoning characteristics of local pixel-level perception and regional structure level to achieve accurate expression of the complex sewage area distribution in remote sensing images, providing highly reliable spatial input basis for subsequent downstream tasks.

[0076] (5) Result output module, 1) Graphics output, In the remote sensing sewage identification task, intuitive graphic output not only helps with result display and manual verification, but also provides visual support for subsequent pollution monitoring and administrative decision-making. However, the original model output is usually presented in the form of a probability map, which is difficult to use directly for interpretation and system integration. Therefore, this module designs an effective graphic output to explicitly superimpose the recognition results on the original remote sensing image to form a visualization result that is both readable and retains image details. Specifically, the first input is the final pollution area prediction probability map output by the aforementioned spatial consistency enhancement and segmentation result optimization module. , where each pixel position Value Indicates the prediction confidence that the pixel belongs to the sewage area.

[0077] In order to extract clear sewage area boundaries, the probability map Perform threshold segmentation to obtain a binary mask map , defined as follows: , in is the probability threshold, which is usually selected through a cross-validation strategy on the training set, aiming to improve the recall ability of the sewage area while maintaining the recognition accuracy. In order to achieve the visualization of the mask map, a color mapping function is constructed. , the binary image Mapped to a three-channel pseudo-color image , for example, using red to highlight the masked area to make the polluted area more visually prominent.

[0078] In the fusion stage, considering the original remote sensing image With rich geographic information background, in order to avoid information shielding, a linear weighted fusion strategy is used to transform the pseudo-color mask map into With the original image According to the transparency factor Fusion, final visualization image The calculation is as follows: , This method highlights the sewage area while maintaining the details of the original image, effectively improving the interpretability. Finally, the image output is saved in a standard image format (such as PNG or GeoTIFF).

[0079] 2) Export of spatial coordinates, In remote sensing wastewater identification tasks, the pollution mask output by the model is typically located in image pixel space and lacks direct geospatial localization capabilities. However, accurate spatial boundary information is essential for downstream pollution monitoring, law enforcement verification, and environmental information system integration. Therefore, based on the model segmentation output, it is necessary to accurately map the pixel-level results to geospatial space, completing the critical transformation from "image recognition" to "geolocation."

[0080] This module generates a binary pollution mask in the aforementioned graphic output (From the probability graph Threshold Based on the data obtained through processing, spatial coordinate conversion and geographic vector boundary derivation are performed. A series of morphological operations, including opening, closing and connected domain analysis, are applied to remove noise and extract the boundaries of contaminated patches. Let the valid patch area finally retained be the set , where each patch Contains several pixel coordinate points , , Consider remote sensing imagery It contains georeference information (such as affine transformation matrix , or RPC projection parameters), can map pixel coordinates to geographic coordinates. Taking the affine model as an example, the transformation formula is as follows: , in, Respectively represent The first of the pollution patches If a nonlinear imaging model such as RPC is used, the corresponding RPC solver should be used to perform spatial coordinate inversion to ensure spatial accuracy.

[0081] Next, for each contaminated patch Corresponding geographic coordinate point set Perform contour sorting. Common methods include Graham scanning, Jarvis March, or convex hull algorithms (such as QuickHull) to construct closed polygon boundaries: , The polygon Contaminated plaque Boundary description in real geographic space. To achieve spatial data management and system calls, all Export to standard spatial formats, such as WKT (Well-Known Text), GeoJSON, or Shapefile, for use in scenarios such as docking with GIS systems, loading into visualization platforms, and pollution tracking and archiving.

[0082] Through the spatial coordinates of this module, the contaminated area identified by the model is derived from the pixel mask in the image domain. Efficiently mapped to vector boundaries in geographic space , realizing the final conversion of remote sensing identification results into geographic decision support data.

[0083] 4. Experimental verification: In order to verify the robustness of the proposed method and the effectiveness of the multi-stage enhancement structure in complex remote sensing scenarios, the following five typical comparison methods are set up: ①UNet: a classic single-scale semantic segmentation model for extracting texture features of remote sensing images; ②PSPNet: a deep image segmentation network with multi-scale perception capabilities, but lacks a structural modeling mechanism; ③GraphUNet: a structural enhancement method that introduces graph neural networks for regional connectivity learning; ④HRNet: a multi-branch fusion architecture that emphasizes high-resolution retention but has no upstream and downstream relationship modeling capabilities; ⑤the proposed method: a complete system integrating multi-resolution residual pyramid, graph structure upstream and downstream modeling, fine-grained edge saliency modeling and result export modules.

[0084] This method uses the same dataset of typical remote sensing pollution images to construct training and test sets, covering a variety of surface types, pollution patterns, and observation conditions, simulating the complex disturbance environments that may occur during actual remote sensing acquisition. Evaluation metrics include: ① Intersection over Union (IoU): reflects the ability to extract the backbone of polluted areas at different scales; ② Boundary Structure Completeness (BCR): assesses the ability to express edge continuity and contour closure of polluted areas; ③ False Alarm Rate (FAR-U): assesses the ability to suppress misjudgments in fuzzy areas; ④ Spatial Projection Error (SPE): assesses the accuracy control of spatial positioning; and ⑤ Inference Time (Time): the processing time required for a single remote sensing image. Higher values ​​for the Intersection over Union (IoU) and BCR indicate better performance, while lower values ​​for the FAR-U, SPE, and Inference Time indicate better performance.

[0085] Table 1 Performance comparison of different methods under five key indicators The experimental results are shown in Table 1. Under the conditions of multi-source disturbance and complex remote sensing background, the traditional method shows obvious performance bottlenecks in recognition accuracy, boundary integrity and system efficiency.

[0086] As a basic convolutional segmentation architecture, UNet has difficulty establishing global perception capabilities when faced with water pollution areas with low texture contrast and blurred boundaries, resulting in a missed detection rate of up to 15.4%. The edges of polluted areas are often mistaken for background areas, and the segmentation results are fragmented and incoherent. PSPNet introduces multi-scale contextual information through the pooling pyramid module, which to a certain extent alleviates the recognition bias caused by deformation. However, it lacks a structural consistency modeling mechanism, and there are still breaks in the edge recognition of planar pollution areas, with a boundary integrity rate of only 74.5%.

[0087] GraphUNet introduces regional connectivity modeling through a graph structure, significantly improving spatial consistency. However, its feature extraction relies primarily on static structural adjacency information and lacks a mechanism to guide upstream and downstream water flow. This results in reduced accuracy in scenarios with drastic changes in watershed boundaries. Furthermore, inference time increases to 1.28 seconds, limiting deployment efficiency. HRNet utilizes multi-branch parallelism to maintain high-resolution features, effectively improving overall recognition accuracy. However, its ability to model spatial topology and causal structures is limited, resulting in ambiguity and overlap in the generation and interpretation of polluted areas.

[0088] By comparison Figure 2 and Figure 3It can be seen that the method of the present invention realizes adaptive perception of pollution forms at different scales by introducing a multi-resolution residual pyramid structure; enhances the spatial topological consistency modeling of water bodies through the upstream and downstream constraint mechanism of the graph structure, effectively reducing the missed detection rate and false alarm rate; with the support of edge saliency modeling and uncertainty suppression mechanism, the model can still output high-quality segmentation masks under complex backgrounds and fuzzy boundaries, and the boundary completeness rate reaches 87.9%; at the same time, the average inference time of the system is only 0.74 seconds, which has good actual deployment efficiency while ensuring accuracy.

[0089] In summary, the method of the present invention has achieved systematic improvements in multiple dimensions such as recognition accuracy, structural integrity and processing efficiency, providing stable and reliable technical support for the identification of sewage areas in remote sensing images, and has significant engineering application value and promotion potential.

[0090] Example 2 This embodiment provides a remote sensing sewage area identification system based on graph structure and multi-stage enhancement, including: The data acquisition module is configured as follows: A computer-readable storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, a remote sensing sewage area identification method based on graph structure and multi-stage enhancement.

[0091] A terminal device includes a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, and the instructions are suitable for being loaded and executed by the processor. A remote sensing sewage area identification method based on graph structure and multi-stage enhancement is described.

[0092] The above are all preferred embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A remote sensing sewage area identification method based on graph structure and multi-stage enhancement, characterized in that: include: Acquisition of remote sensing images; Perform data preprocessing and representation enhancement on the acquired remote sensing images; The enhanced remote sensing images are used to initially screen for significant areas of polluted water, including the construction of anomaly enhancement maps based on local statistical distribution; extraction of pollution candidate maps driven by spatial structure priors; and enhancement of stable regional responses based on a pollution area structural consistency reinforcement mechanism. High-precision segmentation and identification of sewage areas, including multi-resolution residual pyramid structure construction; fine-grained boundary structure modeling and uncertainty suppression; sewage distribution probability map generation and structural consistency optimization; Spatial relationship modeling of sewage areas based on graph attention, including graph structure construction and node relationship initialization; graph attention network modeling and structure enhancement reasoning; spatial consistency enhancement and segmentation result optimization; Output the results.

2. The remote sensing sewage area identification method based on graph structure and multi-stage enhancement according to claim 1 is characterized in that: The data preprocessing and characterization enhancement of the acquired remote sensing images include introducing a spectral consistency normalization processing strategy and performing spectral domain standardization operations on the original images to eliminate spectral drift caused by non-target factors. In order to ensure comparability between bands at the same scale, the spectral vectors of all pixels are first normalized, and the Z-score normalization form is adopted to convert the pixel values ​​of each band into a form with a mean of 0 and a standard deviation of 1. At the same time, in order to enhance the difference expression and feature differentiation ability between the bands, a learnable frequency domain transformation enhancement mechanism is introduced, and the spectral vector is regarded as a one-dimensional signal, and a Fourier transform is performed on it to extract frequency domain features. In order to suppress frequency domain interference and enhance texture expression, a frequency domain enhancement mechanism based on Fourier transform is introduced to achieve the retention of high-frequency structure and the suppression of low-frequency redundancy.

3. The remote sensing sewage area identification method based on graph structure and multi-stage enhancement according to claim 2 is characterized in that: The abnormal enhancement mapping based on local statistical distribution is constructed, including converting the preprocessed and enhanced remote sensing image tensor ,in 、 are the height and width of the image, For the number of bands, an abnormal enhancement mapping method based on the deviation of local statistical distribution is constructed; for each pixel position in the image , extract it in is the center and the side length is Local window , calculate the mean vector of all pixels in the window in the spectral dimension and covariance matrix : ， ， in Indicates the position in the local window The spectral vector of represents the number of pixels in the local window, Represents the mean vector of all pixels in the window in the spectral dimension, represents the covariance matrix.

4. The method for remote sensing sewage area identification based on graph structure and multi-stage enhancement according to claim 3 is characterized in that: The pollution candidate map extraction based on spatial structure prior drive includes pre-processing spectral enhancement image tensor As input, by the prior abnormal response graph Perform weight sorting to obtain , then In the subjectively important band The structure is extracted on the image; the Sobel operator is used to extract the local gradient response of the image in the horizontal and vertical directions, and the overall gradient amplitude map is constructed as the local intensity change response. In order to identify the fuzzy boundary area, the local Laplace transform is introduced to characterize the edge sharpness, and the edge fuzziness scoring function is defined. To highlight the edge abnormality of the contaminated area, the gradient amplitude map is finally and fuzziness score graph Perform weighted fusion to construct a structural saliency map , by setting the significance threshold Binarize the fusion image to extract the pollution candidate area and structural saliency map Expressed as: ， in represents the structural saliency map of all pixels, represents the gradient magnitude map of all pixels, represents the maximum value of the gradient amplitude in the gradient amplitude map, Represents the edge fuzziness scoring function, fusion coefficient This can be obtained by adjusting the parameters of the validation set, and the recommended initial value is 0.

6.

5. The method for remote sensing sewage area identification based on graph structure and multi-stage enhancement according to claim 4 is characterized in that: The structure consistency enhancement mechanism based on the pollution area enhances the stable area response, including the structure candidate map As the initial saliency hint, based on the preprocessed image tensor A structural feature representation is constructed for cross-temporal alignment and consistency scoring. To measure the structural consistency between the current frame and the reference frame within the candidate region, the vector angle cosine similarity is introduced as a consistency metric. Subsequently, to avoid noise propagation caused by directly using the original structure map, only the structural consistency within the candidate region is statistically analyzed to define the final structural consistency enhancement mask.

6. The method for remote sensing sewage area identification based on graph structure and multi-stage enhancement according to claim 5 is characterized in that: The fine-grained boundary structure modeling and uncertainty suppression, including the fusion feature map output by multi-resolution residual pyramid As input, combined with the original spectral image , in the input image Apply Sobel filter to extract spatial gradient intensity and obtain boundary response map , then the boundary response map Intermediate scale features with pyramids Concatenate in the channel dimension and input a lightweight convolutional attention module to extract boundary saliency weights: ， in represents the boundary significance weight, Indicates channel splicing, is a 1×1 convolution operation, is the Sigmoid activation function.

7. The method for remote sensing sewage area identification based on graph structure and multi-stage enhancement according to claim 6 is characterized in that: The sewage distribution probability map generation and structural consistency optimization include the introduction of pollution probability map generation and consistency optimization mechanism based on structural prior map on the basis of completing multi-scale structure fusion and boundary refinement modeling, and outputting the final prediction result with continuous space, smooth boundaries and reasonable structure. Among them, the fusion feature map output based on fine-grained boundary structure modeling , through a convolutional prediction head Generate the probability value P of each pixel belonging to the sewage area, and introduce the consistency regularization term in the training stage The predicted probability map is structurally guided to match the candidate region structure, which is expressed as: ， in represents the consistency regularization term, Indicates the height of the image, Indicates the width of the image, Represents the final structural consistency enhancement map of all pixels, Indicates the probability value of all pixels belonging to the sewage area, Represents a very small positive number used to prevent the denominator from being 0.

8. The method for remote sensing sewage area identification based on graph structure and multi-stage enhancement according to claim 7 is characterized in that: The graph attention network modeling and structure enhancement reasoning include the introduction of graph attention network, which dynamically adjusts the information transmission intensity between adjacent nodes through the multi-head self-attention mechanism, thereby realizing cross-region feature enhancement reasoning, wherein the constructed node set is , the initial feature of each node is , the adjacency weight matrix is , No. Attention head to node The update is expressed as: ， in Indicates the Nodes in the head For Node The attention weight, is the characteristic linear transformation matrix, is the attention parameter, represents vector concatenation, For nodes The set of adjacent nodes, LeakyReLU is a nonlinear activation function used to enhance the model's ability to express negative weights. Representation node The initial characteristics of represents the initial features of each node, Representation node The initial features of the neighboring nodes.

9. The method for remote sensing sewage area identification based on graph structure and multi-stage enhancement according to claim 8 is characterized in that: The spatial consistency enhancement and segmentation result optimization include the introduction of graph enhancement mapping mechanism and structure preservation regularization term to improve the spatial consistency and expression integrity of the final segmentation graph. The node features output by the graph attention network are , which represents the first The semantic representation of the polluted sub-blocks is then used to assign each node feature to its corresponding pixel set through the region inverse mapping operation. , and get the enhanced feature map ; Then enhance the feature map and fusion feature Perform splicing and fusion and input the lightweight decoder layer To generate the final pollution area prediction probability map In order to alleviate the potential risk of feature offset in the graph reasoning process, a structure-preserving regularization term is introduced to measure the final predicted graph Attention input probability map with the original map The spatial structural consistency between them is defined as the weighted KL divergence regularization term: ， in, represents the spatial structure consistency regularization term, represents the probability value of the polluted area, The predicted probability map representing the original contaminated area, Represents the predicted probability map of the final contaminated area, Represents the edge saliency map, defined in the previous module, which is used to enforce the consistency constraints on the boundary area.

10. A remote sensing sewage area identification system based on graph structure and multi-stage enhancement, comprising: The data acquisition module is configured to acquire remote sensing images; The preprocessing module is configured to perform data preprocessing and representation enhancement on the acquired remote sensing image; The initial screening module is configured to perform initial screening of polluted areas in enhanced remote sensing images, including the construction of anomaly enhancement maps based on local statistical distribution; extraction of pollution candidate maps driven by spatial structure priors; and enhancement of stable area responses based on a pollution area structural consistency reinforcement mechanism. The recognition module is configured for high-precision segmentation and recognition of sewage areas, including multi-resolution residual pyramid structure construction; fine-grained boundary structure modeling and uncertainty suppression; sewage distribution probability map generation and structural consistency optimization; The modeling module is configured to model the spatial relationship of sewage areas based on graph attention, including graph structure construction and node relationship initialization; graph attention network modeling and structure enhancement reasoning; spatial consistency enhancement and segmentation result optimization; The output module is configured to output the results.

Images