A remote sensing sewage area identification method and system based on graph structure and multi-stage enhancement
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- YANTAI UNIV
- Filing Date
- 2025-06-30
- Publication Date
- 2026-07-14
Smart Images
Figure CN120726484B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of remote sensing image recognition technology, and in particular to a remote sensing wastewater area identification method and system based on graph structure and multi-stage enhancement. Background Technology
[0002] With the acceleration of industrialization and the continuous expansion of urbanization, water pollution has become increasingly serious, especially in typical areas such as urban river networks, industrial clusters, and near-shore sewage outlets. Frequent sewage discharge, complex composition, and wide spatial distribution pose a serious threat to ecological security and the quality of water environments for human habitation. Remote sensing technology, due to its advantages of large-scale, high-frequency, and non-contact acquisition of water information, has gradually become an important means of sewage identification and monitoring. However, in large-scale remote sensing images, sewage areas often exhibit characteristics such as large scale variations, blurred boundaries, and diverse morphologies, posing a challenge to traditional image analysis methods and urgently requiring more intelligent and automated identification solutions.
[0003] Current 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 level of accuracy in specific water environments, they have significant limitations: they are sensitive to complex backgrounds, struggle to adapt to imaging differences between different remote sensing platforms, and their performance deteriorates drastically, especially when interference is 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, resulting in blurred boundary locations, poor connectivity, and insufficient interpretability.
[0004] Some studies have attempted to introduce deep learning models (such as U-Net and DeepLab) to improve the segmentation accuracy of remote sensing images, training models with large-scale remote sensing data to obtain more robust feature extraction capabilities. However, these methods often focus on the pixel or semantic layer, failing to fully utilize the potential spatial dependencies and structural constraints between polluted areas, and are unable to provide reliable support for irregular boundaries, sparse polluted patches, and other similar situations. Furthermore, due to the scale variations and mixed ground features in 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 not yet formed a mature technical path for water pollution identification. Most methods employ fixed graph structures or simple adjacency mapping, lacking cross-level and cross-regional spatial consistency constraints, making it difficult to form a global collaborative enhancement mechanism. Furthermore, the characteristics of polluted areas, such as blurred boundaries and drastic morphological changes, place higher demands on the robustness of graph structures, making it difficult for traditional graph neural networks to stably capture the potential structural commonalities and differences between regions.
[0006] Therefore, there is an urgent need to propose a remote sensing image wastewater identification method that integrates graph structure modeling capabilities, possesses a spatial collaborative enhancement mechanism, and supports phased expression optimization. This method should be able to accurately model the spatial distribution relationship, boundary features, and salient structure of polluted areas under complex background interference, automatically extract high-precision wastewater area information from the original remote sensing image, and support subsequent multi-level output requirements such as spatial mapping, graphic overlay, and geographic coordinate export, thereby facilitating the intelligent upgrading of environmental monitoring and governance systems. Summary of the Invention
[0007] To address the aforementioned problems in current remote sensing image wastewater identification tasks, such as insufficient identification accuracy, low efficiency in utilizing structural information, weak cross-scale feature expression capabilities, and lack of spatial positioning closed-loop mechanisms, this invention provides a remote sensing wastewater region identification method and system based on graph structure and multi-stage enhancement.
[0008] In a first aspect, the present invention provides a remote sensing wastewater region identification method based on graph structure and multi-stage enhancement, which adopts the following technical solution:
[0009] A remote sensing wastewater region identification method based on graph structure and multi-stage enhancement includes:
[0010] Acquire remote sensing images;
[0011] Data preprocessing and characterization enhancement are performed on the acquired remote sensing images;
[0012] The enhanced remote sensing images are used for initial screening of salient wastewater areas, including: construction of anomaly enhancement maps based on local statistical distribution; extraction of pollution candidate maps driven by spatial structure priors; and enhancement of stable region responses based on pollution region structure consistency enhancement mechanisms.
[0013] High-precision segmentation and identification of wastewater areas, including: construction of multi-resolution residual pyramid structure; fine-grained boundary structure modeling and uncertainty suppression; generation of wastewater distribution probability map and optimization of structural consistency;
[0014] Graph attention-based spatial relationship modeling of wastewater regions includes graph structure construction and node relationship initialization; graph attention network modeling and structure enhancement reasoning; spatial consistency enhancement and segmentation result optimization.
[0015] Output results.
[0016] Furthermore, the data preprocessing and characterization enhancement of the acquired remote sensing images includes introducing a spectral consistency normalization processing strategy. This involves performing spectral domain standardization on the original image to eliminate spectral drift caused by non-target factors. Specifically, to ensure comparability between bands at the same scale, the spectral vectors of all pixels are first normalized using Z-score standardization, converting the pixel values of each band to a form with a mean of 0 and a standard deviation of 1. Simultaneously, to enhance the difference expression and feature discrimination ability between bands, a learnable frequency domain transformation enhancement mechanism is introduced. The spectral vector is treated as a one-dimensional signal, and a Fourier transform is performed to extract frequency domain features. Furthermore, to suppress frequency domain interference and enhance texture expression, a Fourier transform-based frequency domain enhancement mechanism is introduced, achieving the preservation of high-frequency structures and the suppression of low-frequency redundancy.
[0017] Furthermore, the construction of the anomaly enhancement mapping based on local statistical distribution includes processing and enhancing the remote sensing image tensor. ,in , These are the height and width of the image, respectively. For each band, an anomaly enhancement mapping method based on local statistical distribution deviation is constructed; for each pixel position in the image... Extract its in Centered on, with side length as Local window Calculate the mean vector of all pixels within the window in the spectral dimension. Covariance Matrix :
[0018] ,
[0019] ,
[0020] in Indicates the position within a local window spectral vector, This indicates the number of pixels in a local window.
[0021] Furthermore, the pollution candidate map extraction based on spatial structure priors includes enhancing the image tensor with preprocessed spectral data. As input, through the prior anomalous response map To obtain by performing weighted sorting Then to In the subjectively important band Structure extraction is performed on the image; the Sobel operator is used to extract the local gradient response in the horizontal and vertical directions, and an overall gradient magnitude map is constructed as the local intensity change response. To identify blurred boundary regions, a local Laplacian transform is introduced to characterize the edge sharpness, and an edge blur scoring function is defined. To highlight the edge anomalies of the polluted area, the gradient magnitude map is finally used. Ambiguity scoring map Perform weighted fusion to construct a structural saliency map By setting a significance threshold Binarize the fused image to extract contamination candidate regions, and perform a structure saliency map. Represented as:
[0022] ,
[0023] Wherein the fusion coefficient This can be obtained through parameter tuning on the verification set; an initial value of 0.6 is recommended.
[0024] Furthermore, the mechanism for enhancing stable regional response based on the structural consistency of polluted areas includes structural candidate maps. As an initial saliency cue, and also based on the preprocessed image tensor A structural feature representation is constructed for cross-temporal relative alignment and consistency scoring. To measure the structural consistency between the current frame and the reference frame in the candidate region, the cosine similarity of the vector angle is introduced as a consistency metric. Subsequently, to avoid noise propagation caused by directly using the original structural graph, only the structural consistency in the candidate region is statistically analyzed, and the final structural consistency enhancement mask is defined.
[0025] Furthermore, the fine-grained boundary structure modeling and uncertainty suppression include fused feature maps output by multi-resolution residual pyramids. As input, combined with the original spectral image In the input image A Sobel filter is applied to the boundary to extract the spatial gradient intensity, thus obtaining the boundary response map. Then the boundary response map Features at the mesoscale of the pyramids Concatenate along the channel dimension and input a lightweight convolutional attention module to extract boundary saliency weights:
[0026] ,
[0027] in Indicates channel splicing. This is a 1×1 convolution operation. This is the Sigmoid activation function.
[0028] Furthermore, the wastewater distribution probability map generation and structural consistency optimization include, based on the completion of multi-scale structural fusion and boundary refinement modeling, introducing a pollution probability mapping 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. The fusion feature map output based on fine-grained boundary structure modeling is further included. Predicting head via convolution Generate a probability value P for each pixel belonging to the sewage region, and introduce a consistency regularization term during the training phase. The predicted probability map is structurally guided to match the candidate region structure, as shown below:
[0029] ,
[0030] in This represents a consistency regularization term. Indicates the height of the image. Indicates the width of the image. This represents the final structure consistency enhancement map of all pixels. This represents the probability value that all pixels belong to the sewage region. This represents a very small positive number used to prevent the denominator from being zero.
[0031] Furthermore, the graph attention network modeling and structure-enhanced reasoning includes introducing a graph attention network and dynamically adjusting the information transmission intensity between adjacent nodes through a multi-head self-attention mechanism, thereby achieving cross-regional feature-enhanced reasoning. The constructed node set is... Each node has the following initial features: The adjacency weight matrix is , No. Attention head to node The update is represented as:
[0032] ,
[0033] in Indicates the first Node in the head For nodes Attention weights The characteristic linear transformation matrix, For attention parameters, This represents vector concatenation. For nodes The set of adjacent nodes, LeakyReLU is a non-linear activation function used to enhance the model's ability to express negative weights.
[0034] Furthermore, the spatial consistency enhancement and segmentation result optimization include introducing a graph enhancement mapping mechanism and a structure preservation regularization term to improve the spatial consistency and representational integrity of the final segmentation graph. The node features output by the graph attention network are... , indicating the first after structural reasoning The semantic representation of each contaminated sub-block is obtained, and then each node feature is assigned to its corresponding pixel set through a region inverse mapping operation. The enhanced feature map is obtained. Subsequently, the enhanced feature map and fused features will be... Perform splicing and fusion, and input into a lightweight decoder layer. To generate a probability map of the final contaminated area To mitigate the potential feature shift risk during graph reasoning, a structure preservation regularization term is introduced to measure the final predicted graph. Note the input probability graph compared to the original graph. Spatial structural consistency between them is defined by the structural consistency regularization term as a weighted KL divergence form:
[0035] ,
[0036] in, This is the edge saliency map, defined in the previous module, used to strengthen the constraint on the consistency of the boundary region.
[0037] Secondly, a remote sensing wastewater area identification system based on graph structure and multi-stage enhancement includes:
[0038] The data acquisition module is configured to acquire remote sensing images;
[0039] The preprocessing module is configured to perform data preprocessing and characterization enhancement on the acquired remote sensing images;
[0040] The initial screening module is configured to perform initial screening of salient wastewater areas in the enhanced remote sensing image, including: constructing anomaly enhancement maps based on local statistical distribution; extracting pollution candidate maps driven by spatial structure priors; and enhancing the response of stable areas based on the pollution area structure consistency enhancement mechanism.
[0041] The identification module is configured for high-precision segmentation and identification of wastewater areas, including: construction of multi-resolution residual pyramid structure; fine-grained boundary structure modeling and uncertainty suppression; generation of wastewater distribution probability map and optimization of structural consistency.
[0042] The modeling module is configured to perform spatial relationship modeling of wastewater regions based on graph attention, including graph structure construction and node relationship initialization; graph attention network modeling and structure enhancement inference; spatial consistency enhancement and segmentation result optimization.
[0043] The output module is configured to output the results.
[0044] Thirdly, the present invention provides a computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the aforementioned remote sensing wastewater area identification method based on graph structure and multi-stage enhancement.
[0045] Fourthly, the present invention provides a terminal device, including 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, the instructions being adapted to be loaded and executed by the processor to provide a remote sensing wastewater area identification method based on graph structure and multi-stage enhancement.
[0046] In summary, the present invention has the following beneficial technical effects:
[0047] Compared with existing remote sensing image sewage area identification methods, which suffer from weak multi-scale feature extraction capabilities, poor spatial consistency, unstable boundary identification, and lack of geographic registration support, the remote sensing image sewage area identification method based on graph structure modeling and multi-stage enhancement proposed in this invention significantly improves the identification accuracy, boundary representation ability, and geographic projection accuracy of polluted areas in complex remote sensing scenarios.
[0048] First, this invention constructs a multi-resolution residual pyramid structure to fully mine image contextual information at different perception scales, significantly enhancing the model's sensitivity and edge integrity to sewage regions against complex texture backgrounds. Second, by combining fine-grained edge saliency graph modeling and uncertainty suppression mechanisms, edge saliency response weights are introduced when determining boundary regions. Through multi-scale perturbation perception and confidence constraints, background interference and misjudgments of blurred regions are effectively suppressed. Third, through graph structure modeling and upstream / downstream information constraint mechanisms, the spatial topological relationships of polluted water bodies are modeled and corrected, effectively enhancing the consistent expression of regional structures. Finally, an output module is designed to accurately map the recognition result mask image to the image and geospatial coordinate system, providing data support for practical pollution control and regulatory applications.
[0049] In a typical remote sensing pollution image test set, the average identification accuracy (IoU) of the method of this invention for polluted areas was improved from 81.3% of the traditional method to 92.7%; the boundary structure integrity score was improved from 72.6% to 87.9%, significantly improving the representation of area 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 overall average inference time of the system is 0.74 seconds per image, demonstrating good timeliness and deployment stability. With advantages such as high identification accuracy, strong boundary stability, accurate geographic mapping, and excellent system response efficiency, this method achieves highly reliable identification of sewage areas in remote sensing scenarios, demonstrating significant engineering practical value and broad application prospects. Attached Figure Description
[0050] Figure 1 This is a schematic diagram of a remote sensing wastewater area identification method based on graph structure and multi-stage enhancement according to Embodiment 1 of the present invention.
[0051] Figure 2 This is a comparison chart showing the IoU and BCR indicators of the various methods in Embodiment 1 of the present invention.
[0052] Figure 3 This is a comparison chart showing the results of the various methods in Embodiment 1 of the present invention on FAR-U, SPE, and Time metrics. Detailed Implementation
[0053] The present invention will be further described in detail below with reference to the accompanying drawings.
[0054] Example 1
[0055] Reference Figure 1 This embodiment of a remote sensing wastewater region identification method based on graph structure and multi-stage enhancement includes:
[0056] (1) Data preprocessing and characterization enhancement module,
[0057] 1) Spectral consistency normalization processing,
[0058] In the actual acquisition of remote sensing images, due to limitations in imaging sensor performance, differences in imaging angles, changes in atmospheric conditions, and variations in surface reflectivity, remote sensing images of the same area acquired at different times or with different equipment often exhibit inconsistent spectral responses. This spectral difference leads to significant variations in the spectral representation of the same ground feature in different images, thus interfering with the subsequent identification, modeling, and analysis of wastewater areas, and reducing the stability and generalization ability of the model.
[0059] To address the aforementioned issues, this module introduces a spectral consistency normalization processing strategy. This strategy performs spectral domain standardization on the original image to eliminate spectral drift caused by non-target factors and improve the contrast consistency and stability between different bands.
[0060] Suppose the original remote sensing image is represented as a three-dimensional tensor:
[0061] ,
[0062] in, , These represent the height and width of the image, respectively. Indicates the number of bands. For the th band in the image... For each pixel, its corresponding multispectral vector can be represented as:
[0063] ,
[0064] To ensure comparability between bands at the same scale, the spectral vectors of all pixels are first normalized. Z-score normalization is used to convert the pixel values of each band into a form with a mean of 0 and a standard deviation of 1.
[0065] ,
[0066] in, and They represent the first The mean and standard deviation of each band in the entire image This represents the normalized pixel value.
[0067] Meanwhile, to further enhance the ability to express differences and distinguish features among different bands, a learnable frequency domain transform enhancement mechanism is introduced. The spectral vector is treated as a one-dimensional signal, and its Fourier transform is performed to extract frequency domain features.
[0068] ,
[0069] in This indicates the Fourier transform operation. This represents the corresponding frequency domain response. To suppress high-frequency noise while preserving the main spectral structure information, a frequency domain weighted mask is designed. Used to enhance or weaken the corresponding frequency band:
[0070] ,
[0071] Finally, the enhanced time-domain representation is recovered using inverse Fourier transform:
[0072] ,
[0073] in This represents the enhanced time-domain representation. This indicates the inverse Fourier transform operation. The corresponding frequency band after enhancement or attenuation, This indicates the Fourier transform operation. This represents all pixel values after normalization. This represents the frequency domain weight mask.
[0074] After the above processing, the spectral vectors of all pixels are mapped to a feature space with a more stable spectral structure and greater robustness to anomalous bands, thus providing a more reliable input representation for subsequent region identification and modeling.
[0075] 2) Frequency domain shadow interference suppression and texture enhancement.
[0076] Remote sensing images are often affected by factors such as shadows and strong reflections during the natural imaging process, especially in densely built-up areas such as cities and industrial zones. The shadows cast by large buildings or three-dimensional structures at different solar altitude angles can significantly interfere with the spectral representation and spatial texture distribution of the image, leading to misidentification of polluted areas or missed detections. In addition, local texture distortion caused by sensor response and differences in ground material in remote sensing images may also obscure the discernible features of sewage areas at the spatial structural level.
[0077] 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, so as to preserve the high-frequency structure and suppress the low-frequency redundancy.
[0078] Specifically, the normalized image obtained after the previous spectral consistency normalization process is as follows:
[0079] ,
[0080] in, , , These represent the image's height, width, and number of spectral channels, respectively. We first analyze each spectral channel... Perform a two-dimensional Fourier transform to obtain its spectral representation:
[0081] ,
[0082] in Indicates the first Perform a two-dimensional Fourier transform operation on each pixel. This indicates that a two-dimensional Fourier transform operation is performed on each spectral channel. Indicates the height of the image. Indicates the width of the image.
[0083] After transforming all channels, the overall spectrum expression can be obtained:
[0084] ,
[0085] in Indicates the height of the image. Indicates the width of the image. Indicates the number of spectral channels. This represents the result of the Fourier transform operation on each channel. This represents the overall spectrum after all channels have been transformed.
[0086] In the frequency spectrum, the low-frequency components of an image mainly correspond to overall brightness and large-scale regional gradations, which are often significantly affected by shadow interference; while the high-frequency components carry structural information such as image edges and textures, and are important criteria for identifying sewage areas. Therefore, constructing a frequency domain mask is crucial. To retain only the high-frequency region:
[0087]
[0088] in Let be the radius of the frequency domain high-pass filter. The suppression result obtained after applying the frequency domain mask is:
[0089] ,
[0090] in Represents the frequency domain mask. This represents the image after spectral uniformity normalization. This indicates the Fourier transform operation. This represents the suppression result obtained after applying a frequency domain mask.
[0091] Performing an inverse Fourier transform on the high-frequency image in the frequency domain yields the enhanced image representation:
[0092] ,
[0093] in Indicates the height of the image. Indicates the width of the image. Indicates the number of spectral channels. This indicates the inverse Fourier transform operation. This indicates the enhanced image representation.
[0094] To further enhance the distribution characteristics of texture perception, we introduce a texture enhancement factor. The original image and the frequency domain enhancement image are weighted and fused, and the final enhanced image is expressed as follows:
[0095] ,
[0096] in This ensures consistency between inputs and outputs across modules, based on the original input from the "spectral consistency normalization processing". 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 areas, facilitating accurate segmentation of subsequent salient regions.
[0097] 3) Dynamic band selection and information redundancy suppression,
[0098] Remote sensing images, especially hyperspectral images, contain a large number of bands (up to hundreds). While this provides rich spectral information for fine-grained identification, it also brings serious problems of the "curse of dimensionality" and "redundancy interference." On the one hand, many bands are highly correlated, resulting in redundant information and inefficient feature representation. 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" that introduce interference into the model and affect the accurate identification of wastewater areas.
[0099] To address this, this module proposes a dynamic band selection mechanism based on information contribution to maximize effective information and suppress redundant bands, thereby 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 the enhanced features in the frequency domain, resulting in a band selection method that is more adaptable to the task objectives.
[0100] The image tensor after the previous stage of Fourier texture enhancement processing is:
[0101] ,
[0102] in, , Represents the spatial dimension of the image. Indicates the number of bands. Indicates the first Two-dimensional feature map of each band.
[0103] Then define the information entropy for each band. The formula for its independent information contribution is as follows:
[0104] ,
[0105] in, Indicates band Upper The probability density corresponding to each pixel value can be obtained through histogram normalization estimation. This represents the total number of intervals for pixel value discretization. Higher information entropy indicates richer discriminative information contained in the band, making it more valuable for selection.
[0106] To further measure the degree of mutual information redundancy between bands, a mutual information matrix is introduced. :
[0107] ,
[0108] in Indicates band With band Upper pixel value combination The joint probability, and These represent marginal probabilities.
[0109] Next, the spectral selection fractional function is introduced. As the final band evaluation standard:
[0110] ,
[0111] This score combines the information contribution of the band itself with the redundancy of other bands; a higher value indicates that the band is more worthwhile to retain. This is achieved by analyzing all... of Sort and select the previous ones. The optimized band set consists of several bands:
[0112] ,
[0113] The final optimized image tensor used for subsequent processing is:
[0114] ,
[0115] It is worth noting that Fourier enhancement produces It has higher frequency texture and edge sharpness, thus enabling the above process to It is more distinctive, avoids the misselection of low-frequency redundant bands, and improves the targeting and robustness of band selection.
[0116] This dynamic band selection mechanism not only effectively compresses data dimensions and improves model training and inference efficiency, but also eliminates noisy bands and strengthens effective bands, thus laying a high-quality data foundation for subsequent spatial saliency analysis and fine segmentation of wastewater areas.
[0117] (2) Primary screening module for significant wastewater areas,
[0118] 1) Construction of anomaly enhancement mapping based on local statistical distribution.
[0119] In remote sensing image analysis tasks, water bodies and polluted wastewater areas often exhibit significant differences in spectral and textural features. Especially at the local scale, polluted areas, due to the influence of chemical emissions or structural disturbances, typically display the following anomalies: spectral response shifts, fragmented texture structures, and anomalies in high or low reflectance intensity. These anomalies are easily diluted by traditional global statistical features; therefore, it is necessary to introduce spatially aware local modeling methods to perform prior enhancement on potentially polluted areas, thereby improving the sensitivity of subsequent segmentation modules to significantly polluted regions.
[0120] Therefore, this module is based on the remote sensing image tensors that have been preprocessed and enhanced in the previous stage. ,in , These are the height and width of the image, respectively. For the number of bands, an anomaly enhancement mapping method based on the deviation of local statistical distribution was constructed.
[0121] For the position of each pixel in the image Extract its in Centered on, with side length as Local window Calculate the mean vector of all pixels within the window in the spectral dimension. Covariance Matrix :
[0122] ,
[0123] ,
[0124] in Indicates the position within a local window spectral vector, This indicates the number of pixels in a local window.
[0125] Subsequently, the Mahalanobis distance is used to measure the deviation between the current pixel and its local statistical distribution:
[0126] ,
[0127] in, Represents pixels The degree of anomaly relative to the statistical characteristics of its neighborhood. The larger the value, the less the point deviates from the statistical distribution of its local environment, and the more likely it is to belong to a potentially polluted area.
[0128] Finally, the Mahalanobis distance values of all pixels in the entire image are used to construct a pollution anomaly response map. And perform normalization:
[0129] ,
[0130] Normalized anomaly graph It can serve as an important prior information channel for subsequent wastewater segmentation modules, enhancing the network's ability to perceive potentially polluted areas and providing key support for refined segmentation and boundary modeling.
[0131] 2) Spatial structure prior-driven pollution candidate map extraction,
[0132] Wastewater areas in remote sensing images not only exhibit anomalous reflectance characteristics in spectral space but also show significant distributional differences at the spatial structure level. Compared to clean water bodies, which typically exhibit clear boundaries and smooth textures, polluted water bodies tend to have inconsistencies in local structure, such as blurred boundaries, irregular shapes, and dramatic fluctuations in internal intensity. These spatial structural differences reflect the disturbances that pollution processes cause to the continuity and surface state of water bodies, and serve as important evidence for constructing spatial priors.
[0133] To further extract structurally significant contamination candidate regions, this module designs a spatial saliency construction method that integrates image gradient response and edge blurring score to reveal the structural anomalies of contamination regions at the spatial distribution level. First, the spectral enhancement image tensor output by the preprocessing module is used... As input, through the prior anomalous response map To obtain by performing weighted sorting Then to In the subjectively important band Structure extraction is performed on the image. Let the corresponding two-dimensional image be... .
[0134] The classic Sobel operator is used to extract the local gradient responses of an image in the horizontal and vertical directions, defined as:
[0135] ,
[0136] Then, an overall gradient magnitude map is constructed as a response to local intensity changes:
[0137] ,
[0138] The larger the gradient magnitude, the more significant the texture changes or edge abrupt changes in the local area, which may correspond to areas of pollution disturbance.
[0139] To further identify regions with blurred boundaries, a local Laplacian transform is introduced to characterize edge sharpness. Let the image be located at... The response at that location is:
[0140] ,
[0141] in This represents the discrete Laplacian operator. Define the edge ambiguity scoring function. for:
[0142] ,
[0143] The rating is The normalization factor is applied within a certain range; a larger value indicates a more blurred boundary and greater uncertainty. This factor can effectively suppress interference from smooth edges in the background water, highlighting the edge anomalies of polluted areas.
[0144] To construct the spatial structure saliency fusion map, the gradient magnitude map is used. Ambiguity scoring map Perform weighted fusion to construct a structural saliency map :
[0145] ,
[0146] Wherein the fusion coefficient This can be obtained through validation set parameter tuning, with a recommended initial value of 0.6. Subsequently, a binary mask is used to extract contaminated candidate regions. A significance threshold is then set. Binarize the fused image to extract candidate contamination regions:
[0147] ,
[0148] Finally, the candidate mask is obtained. As a regional concern constraint for subsequent wastewater fine segmentation and upstream and downstream modeling modules, it reduces interference from redundant backgrounds and improves processing efficiency and accuracy.
[0149] 3) Mechanism for enhancing structural consistency in polluted areas.
[0150] In multi-temporal observations of remote sensing images, the spectral and spatial structural characteristics of contaminated areas may change significantly with time, shooting angle, weather conditions, and other factors. This can lead to structural shifts, blurred boundaries, or inconsistent responses in the same contaminated area at different time points. Without modeling and correction, this can easily result in false positives or false negatives. In particular, isolated outliers or background disturbances can interfere with the consistent representation of area boundaries, reducing model stability.
[0151] To improve the structural stability and continuity of the initial candidate regions in the temporal dimension, this module proposes a temporal consistency enhancement mechanism based on structural feature similarity. By performing structural alignment operations between multi-temporal images, the response of stable regions is enhanced and the influence of structurally drifting regions is suppressed.
[0152] This module contains the structural candidate diagram output by the previous module. As an initial saliency cue, and also based on the preprocessed image tensor Construct structural feature representations for cross-temporal relative homogeneity and consistency scoring. Let the current time phase be... With reference phase The candidate structures are as follows:
[0153] ,
[0154] Then, the input image tensor is processed by the structural feature extraction function. Obtain structural characterization diagram:
[0155] ,
[0156] in Dimensions representing structural features, such as Sobel edge maps, Laplacian maps, or texture orientation features.
[0157] To measure the structural consistency between the current frame and the reference frame within the candidate region, the cosine similarity of the vector angle is introduced as a consistency metric:
[0158] ,
[0159] in Indicates the pixel position in the image. Represents the vector dot product. To prevent the use of small constants related to division by zero, and to avoid noise propagation from directly using the original structure diagram, we then statistically analyze the structural consistency only within the candidate regions, defining a final structural consistency enhancement mask:
[0160] ,
[0161] If a region has low structural similarity (e.g., texture abrupt changes due to specular reflection, sensor noise, or cloud occlusion), its saliency response will be attenuated. Conversely, if the region's structure remains consistent across two frames, its response will be preserved or enhanced. The final structural consistency enhancement map obtained after the above processing is as follows:
[0162] ,
[0163] This will serve as input prompts for subsequent fine segmentation and pollution attribute modeling modules, guiding the fine segmentation of structurally continuous regions.
[0164] (3) High-precision segmentation and recognition module for wastewater areas.
[0165] 1) Construction of multi-resolution residual pyramid structure,
[0166] In remote sensing images, sewage 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, thus limiting high-precision segmentation performance. To address this, this module introduces a multi-resolution residual pyramid (MRP) structure to enhance the network's ability to represent sewage areas at different spatial scales and achieve fusion perception of structural details and contextual semantic information.
[0167] The input to this module is the structural consistency enhancement obtained from the initial screening module. Figure 2 dimensional tensor ,in and These are the height and width of the image, respectively. We first... Constructing multi-level pyramid feature sequences , of which Layered pyramid diagram By downsampling factor Generated from the original image, i.e.:
[0168] ,
[0169] To enhance the nonlinear representation capability at each scale, a residual encoding block (REB) is introduced to perform residual enhancement on the features of each pyramid level. Let the... Layer input features are Then its output after passing through the residual module is:
[0170] ,
[0171] in For learnable convolutional kernels, This represents the convolution operation. It is the ReLU activation function. The sigmoid function is used, and residual connections are employed to preserve local details of the original input. Subsequently, residual feature maps at different scales are fused from the bottom up to construct a unified high-dimensional multi-scale structural representation. :
[0172] ,
[0173] Among all the smaller-scale feature maps All channels are upsampled to their original resolution before being stitched together. The final output is a fused feature map. It possesses rich hierarchical structural information, which can provide feature support with strong discriminative power and spatial consistency for subsequent processing.
[0174] 2) Fine-grained boundary structure modeling and uncertainty suppression.
[0175] In remote sensing scenarios, the edges of sewage areas often exhibit high grayscale transitions and local structural ambiguity. Especially when affected by interference from lighting, wind, and waves, the boundary between the polluted area and the background water body becomes blurred, causing traditional segmentation methods to easily produce errors such as "false contours" or "fragmentation" in the edge areas.
[0176] To enhance the model's ability to identify structures in boundary regions, this module proposes a fine-grained structure modeling mechanism based on gradient perception and boundary attention guidance, which dynamically suppresses prediction uncertainty in edge transition regions while preserving spatial continuity.
[0177] This module uses the fused feature map output from the previous stage multi-resolution residual pyramid module. As input, combined with the original spectral image Perform the following operations:
[0178] First, in the input image A Sobel filter is applied to the boundary to extract the spatial gradient intensity, thus obtaining the boundary response map. :
[0179] ,
[0180] in, , The horizontal and vertical convolution kernels of the Sobel operator. This represents a two-dimensional convolution operation. The boundary response map is then used. Features at the mesoscale of the pyramids Concatenate along the channel dimension and input a lightweight convolutional attention module to extract boundary saliency weights:
[0181] ,
[0182] in Indicates channel splicing. This is a 1×1 convolution operation. The activation function is Sigmoid. The result is... Each pixel is assigned a weight representing the saliency of its boundary structure. Then, the original fused feature map is weighted using a boundary attention map to enhance the discriminative features of the boundary regions and suppress response perturbations in areas of structural uncertainty.
[0183] ,
[0184] in This indicates element-wise multiplication. This weighted strategy essentially implements a mechanism of boundary region enhancement (attention boost) and non-boundary region suppression (residual skip), making the model more accurate in identifying the edges of contaminated regions.
[0185] 3) Generation and structural consistency optimization of wastewater distribution probability maps.
[0186] Wastewater areas in remote sensing images often exhibit spatial discontinuities, blurred boundaries, and local abrupt changes. Directly performing pixel-by-pixel discrimination based on feature maps can easily lead to the expansion of artifact regions or fragmentation of the target area. To address this, this module, after completing multi-scale structural fusion and boundary refinement modeling, introduces a pollution probability mapping generation and consistency optimization mechanism based on structural prior maps, outputting a final prediction result that is spatially continuous, has smooth boundaries, and a reasonable structure.
[0187] Fusion feature map output based on fine-grained boundary structure modeling It will use a lightweight convolutional prediction head Generate the probability value for each pixel that belongs to the sewage region:
[0188] ,
[0189] in Represents pixels Confidence level that it belongs to the polluted area It can be composed of a set of consecutive 1×1 convolutions and sigmoid activations. This takes into account that structurally stable candidate region masks have already been generated during the initial screening stage. We introduce a consistency regularization term during the training phase. The predicted probability map is structure-guided to match the candidate region structure:
[0190] ,
[0191] Furthermore, to alleviate the ambiguity and instability of boundary region prediction, the boundary attention weight map generated in the previous module is integrated. Dynamic confidence adjustments are made to the preliminary prediction results:
[0192] ,
[0193] in This represents a mild spatial Gaussian smoothing operator, used to suppress local noise and improve edge consistency. The final generated structure-optimized contamination probability map... This will serve as the input for the next stage module, "Spatial Relationship Modeling of Wastewater Regions Based on Graph Attention," providing a high-confidence semantic foundation for spatial mapping of polluted areas and upstream-downstream interactions.
[0194] (4) Graph attention-based modeling module for spatial relationships in wastewater areas.
[0195] 1) Graph structure construction and node relationship initialization,
[0196] In remote sensing images, wastewater areas often exhibit strong spatial continuity, significant local clustering, and complex structural features. This distribution is constrained by non-Euclidean spatial factors such as topographic slope, water flow direction, and anthropogenic discharge paths, and is not random or independent. Therefore, traditional independent prediction methods based on pixels or local windows struggle to capture the deep structural dependencies between regions, and are prone to misjudgments such as prediction breaks, boundary jumps, and hole expansion.
[0197] 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. Polluted areas are divided into structured nodes, and their spatial and semantic relationships are represented using graph structures, thereby capturing non-local inter-regional dependencies.
[0198] The pollution probability map output by the previous module As an input, a probability threshold is first set. Obtain the pseudo-pollution mask image :
[0199] ,
[0200] Subsequently, based on Perform connected component extraction, defining each contaminated connected component as a graph node, and construct a node set. Each node For a given pseudo-contaminated sub-region, its initial feature vector is obtained by fusing feature maps within that region. The pixel mean was calculated as follows:
[0201] ,
[0202] in, Represents a node The corresponding set of pixel coordinates. Secondly, the edge weights of the graph are determined by the spatial location and semantic feature similarity between nodes; therefore, the adjacency weight matrix of the graph is defined. :
[0203] ,
[0204] in, Represents a node Spatial centroid coordinates, 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. The node initialization and adjacency construction method provides structural priors and feature foundations for subsequent graph attention network modeling and structure-enhanced inference.
[0205] 2) Graph attention network modeling and structure-enhanced reasoning.
[0206] Although the aforementioned graph structure has effectively established spatial and semantic connections between wastewater areas, the dependencies between nodes still need to be further modeled through explicit information exchange mechanisms. Traditional graph convolution methods rely on static adjacency structures for weight allocation, making it difficult to fully exploit semantic differences and boundary uncertainties across regions. In remote sensing images, different wastewater areas may exhibit semantically similar structural features despite being geographically distant due to factors such as terrain occlusion and changes in illumination. Therefore, it is urgent to introduce mechanisms with adaptive mapping capabilities to enhance the consistent representation of structures.
[0207] To address this, this module introduces a Graph Attention Network (GAT), which dynamically adjusts the information transmission intensity between neighboring nodes through a multi-head self-attention mechanism, thereby enabling cross-regional feature-enhanced inference.
[0208] The previously constructed node set is Each node has the following initial features: The adjacency weight matrix is , No. Attention head to node The update is represented as:
[0209] ,
[0210] in Indicates the first Node in the head For nodes Attention weights The characteristic linear transformation matrix, For attention parameters, This represents vector concatenation. For nodes The set of adjacent nodes, LeakyReLU is a non-linear activation function used to enhance the model's ability to express negative weights. Then, all... The outputs of each attention head are concatenated to obtain the updated node representation:
[0211] ,
[0212] The aforementioned 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. This alleviates problems such as breaks and missed detections, and enhances the ability to model cross-regional structural consistency. The final node features... This will serve as the structurally enhanced representation, providing a structure-aware semantic foundation for subsequent full graph integration and mask graph remapping.
[0213] 3) Enhanced spatial consistency and optimized segmentation results.
[0214] While graph attention mechanisms can capture latent semantic relationships in distant regions at the structural level, their operational units are still primarily region-level nodes, making it difficult to directly constrain the consistency of fine-grained spatial pixel distribution in the original image. Therefore, to achieve effective fusion between graph structure modeling results and the spatial representation of the original image, this module introduces a graph enhancement mapping mechanism and a structure preservation regularization term to improve the spatial consistency and representational integrity of the final segmentation graph.
[0215] Note the node characteristics of the network output in the diagram. , indicating the first after structural reasoning The semantic representation of each contaminated sub-block is obtained, and then each node feature is assigned to its corresponding pixel set through a region inverse mapping operation. The enhanced feature map is obtained. :
[0216] ,
[0217] Subsequently, the feature maps will be enhanced and the features will be fused. Perform splicing and fusion, and input into a lightweight decoder layer. To generate a probability map of the final contaminated area :
[0218] ,
[0219] in This represents the channel-dimensional concatenation operation. To mitigate the potential feature shift risk during graph inference, a structure-preserving regularization term is further introduced to measure the final predicted graph. Note the input probability graph compared to the original graph. Spatial structural consistency between them. The structural consistency regularization term is defined as a weighted KL divergence form:
[0220] ,
[0221] in, This is the edge saliency map, defined in the previous module, used to strengthen the constraint on the consistency of boundary regions. The final output is a probability map. It integrates local pixel-level perception with collaborative reasoning at the regional structure level, enabling accurate representation of complex sewage area distribution in remote sensing images, and providing highly reliable spatial input for subsequent downstream tasks.
[0222] (5) Result output module,
[0223] 1) Graphics output,
[0224] In remote sensing wastewater identification tasks, intuitive graphical output not only facilitates result presentation 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 graphical output that explicitly overlays the identification results onto the original remote sensing image, forming a visualization 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. The position of each pixel value This indicates the prediction confidence that the pixel belongs to the sewage area.
[0225] To extract clear boundaries of wastewater areas, it is necessary to analyze the probability map. Threshold segmentation is performed to obtain a binary mask image. The definition is as follows:
[0226] ,
[0227] in The probability threshold is typically selected using a cross-validation strategy on the training set, aiming to improve recall of wastewater areas while maintaining recognition accuracy. To achieve visualization of the mask image, a color mapping function is constructed. , binary image Mapped to a three-channel pseudo-color image For example, using red highlights to indicate masked areas makes contaminated areas more visually noticeable.
[0228] During the fusion stage, the original remote sensing images are taken into account. With rich geographic information background, to avoid information masking, a linear weighted fusion strategy is adopted to transform the pseudo-color mask image. With the original image According to transparency factor Fusion, final visualized image The calculation is as follows:
[0229] ,
[0230] This method highlights polluted areas while preserving the original image's details, effectively improving interpretability. General settings. This achieves good visual results. Finally, the image output is saved in a standard image format (such as PNG or GeoTIFF).
[0231] 2) Spatial coordinate export,
[0232] In remote sensing wastewater identification tasks, the pollution mask maps output by the models are typically located in the image pixel space, lacking direct geospatial positioning 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 crucial transformation from "image recognition" to "geolocation."
[0233] This module generates a binary contamination mask based on the aforementioned graphic output. (from probability graph) Threshold Based on the obtained data, spatial coordinate transformation and geographic vector boundary export operations are performed. First, for... A series of morphological operations, including opening, closing, and connected component analysis, are applied to remove noise and extract contaminated plaque boundaries. Let the final retained effective plaque regions be denoted as a set. Each of the patches Contains several pixel coordinates ,
[0234] ,
[0235] Considering remote sensing imagery It contains georeferenced information (such as affine transformation matrices). (or RPC projection parameters), which can map pixel coordinates to geographic coordinates. Taking the affine model as an example, its transformation formula is as follows:
[0236] ,
[0237] in, They represent the first The first of the pollution patches The latitude and longitude of each pixel are calculated. If a nonlinear imaging model such as RPC is used, the corresponding RPC solver is used to perform spatial coordinate inversion to ensure spatial accuracy.
[0238] Next, for each contaminated patch Corresponding set of geographic coordinates Perform contour sorting. Common methods include Graham scan, Jarvis March, or convex hull algorithms (such as QuickHull) to construct closed polygon boundaries.
[0239] ,
[0240] The polygon That is, pollution patches Boundary description in real geospatial space. To achieve spatial data management and system access, ultimately all... Export to standard spatial formats, such as WKT (Well-Known Text), GeoJSON, or Shapefile, for use in scenarios such as interfacing with GIS systems, loading visualization platforms, and creating pollution tracking records.
[0241] By deriving the spatial coordinates from this module, the contaminated areas identified by the model are derived from the pixel mask in the image domain. Effectively mapped to vector boundaries in geospatial space This enables the final conversion of remote sensing identification results into geographic decision support data.
[0242] 4. Experimental verification:
[0243] To verify the robustness of the method of this invention in complex remote sensing scenarios and the effectiveness of the multi-stage enhancement structure, the following five typical comparison methods were set up: ①UNet: a classic single-scale semantic segmentation model used to extract texture features of remote sensing images; ②PSPNet: a deep image segmentation network with multi-scale perception capabilities, but lacking a structure modeling mechanism; ③GraphUNet: a structure enhancement method that introduces graph neural networks for regional connectivity learning; ④HRNet: a multi-branch fusion architecture that emphasizes high-resolution preservation, but lacks the ability to model upstream and downstream relationships; ⑤The method of this invention: a complete system integrating multi-resolution residual pyramids, graph structure upstream and downstream modeling, fine-grained edge saliency modeling, and result export modules.
[0244] This invention uses the same typical remote sensing pollution image dataset to construct both the training and test sets, covering various land surface types, pollution morphologies, and observation conditions to simulate the complex disturbance environment that may occur during actual remote sensing acquisition. Evaluation metrics include: ① Intersection ratio (IoU): reflecting the ability to extract the backbone of polluted areas at different scales; ② Boundary structure integrity score (BCR): assessing the ability to express the continuity and contour closure of wastewater area edges; ③ False alarm rate in uncertain regions (FAR-U): assessing the ability to suppress false alarms in ambiguous areas; ④ Spatial projection error (SPE): assessing the accuracy control capability in spatial positioning; ⑤ Inference time: the processing time for a single remote sensing image. Higher values for IoU and BCR indicate better performance; lower values for FAR-U, SPE, and Inference time indicate better performance.
[0245] Table 1. Performance comparison of different methods under five key indicators
[0246]
[0247] The experimental results are shown in Table 1. Under the conditions of multi-source disturbance and complex remote sensing background, traditional methods show obvious performance bottlenecks in terms of identification accuracy, boundary integrity and system efficiency.
[0248] UNet, as a basic convolutional segmentation architecture, struggles to establish global perception capabilities when facing water pollution areas with low texture contrast and blurred boundaries, resulting in a false negative rate as high as 15.4%. It often misidentifies the edges of pollution areas as background areas, leading to fragmented and discontinuous segmentation results. PSPNet introduces multi-scale contextual information through pooling pyramid modules, which alleviates the recognition bias caused by deformation to some extent. However, it lacks a structural consistency modeling mechanism, and there are still breakage phenomena in the edge recognition of planar pollution areas, with a boundary integrity rate of only 74.5%.
[0249] GraphUNet introduces regional connectivity modeling through graph structure, significantly improving spatial consistency. However, its feature extraction mainly relies on static structural adjacency information and lacks a mechanism to guide upstream and downstream water flow, leading to decreased accuracy in scenarios with drastic changes in watershed boundaries. Simultaneously, 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 structure is limited, resulting in ambiguity and overlap issues in the interpretation of polluted area generation.
[0250] In comparison Figure 2 and Figure 3As can be seen, the method of this invention achieves adaptive perception of pollution morphology at different scales by introducing a multi-resolution residual pyramid structure; it enhances the spatial topological consistency modeling of water bodies through the upstream and downstream constraint mechanism of graph structure, effectively reducing the false alarm rate and the missed detection 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, with a boundary integrity rate of 87.9%; at the same time, the average inference time of the system is only 0.74 seconds, which has good practical deployment efficiency while ensuring accuracy.
[0251] In summary, the method of this invention has achieved a systematic improvement in multiple dimensions such as identification 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.
[0252] Example 2
[0253] This embodiment provides a remote sensing wastewater area identification system based on graph structure and multi-stage enhancement, including:
[0254] The data acquisition module is configured as follows:
[0255] A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device, the aforementioned remote sensing wastewater area identification method based on graph structure and multi-stage enhancement.
[0256] A terminal device includes a processor and a computer-readable storage medium, the processor being configured to implement various instructions; the computer-readable storage medium being configured to store multiple instructions adapted for loading and execution by the processor of the aforementioned remote sensing wastewater area identification method based on graph structure and multi-stage enhancement.
[0257] 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, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.
Claims
1. A remote sensing wastewater region identification method based on graph structure and multi-stage enhancement, characterized in that, include: Acquire remote sensing images; Data preprocessing and characterization enhancement are performed on the acquired remote sensing images; The enhanced remote sensing images are used for initial screening of salient wastewater areas, including: construction of anomaly enhancement maps based on local statistical distribution; extraction of pollution candidate maps driven by spatial structure priors; and enhancement of stable region responses based on pollution region structure consistency enhancement mechanisms. High-precision segmentation and identification of wastewater areas, including: construction of multi-resolution residual pyramid structure; fine-grained boundary structure modeling and uncertainty suppression; generation of wastewater distribution probability map and optimization of structural consistency; Graph attention-based spatial relationship modeling of wastewater regions includes graph structure construction and node relationship initialization; graph attention network modeling and structure enhancement reasoning; spatial consistency enhancement and segmentation result optimization. Output results; The contamination candidate map extraction based on spatial structure priors includes enhancing the image tensor with preprocessed spectral data. As input, through the prior anomalous response map To obtain by performing weighted sorting Then to In the subjectively important band Structure extraction is performed on the image; the Sobel operator is used to extract the local gradient response in the horizontal and vertical directions, and an overall gradient magnitude map is constructed as the local intensity change response. To identify blurred boundary regions, a local Laplacian transform is introduced to characterize the edge sharpness, and an edge blur scoring function is defined. To highlight the edge anomalies of the polluted area, the gradient magnitude map is finally used. Ambiguity scoring map Perform weighted fusion to construct a structural saliency map By setting a significance threshold Binarize the fused image to extract contamination candidate regions, and perform a structure saliency map. Represented as: ,in Represents the structural saliency map of all pixels. A gradient magnitude map representing all pixels. This represents the maximum value of the gradient magnitude in the gradient magnitude graph. This represents the edge ambiguity scoring function, and the fusion coefficient. This can be obtained through parameter tuning on the verification set; an initial value of 0.6 is recommended. The fine-grained boundary structure modeling and uncertainty suppression include fused feature maps output from multi-resolution residual pyramids. As input, combined with the original spectral image In the input image A Sobel filter is applied to the boundary to extract the spatial gradient intensity, thus obtaining the boundary response map. Then the boundary response map Features at the mesoscale of the pyramids Concatenate along the channel dimension and input a lightweight convolutional attention module to extract boundary saliency weights: , in Represents the significance weight of the boundary. Indicates channel splicing. This is a 1×1 convolution operation. Use the Sigmoid activation function; The wastewater distribution probability map generation and structural consistency optimization include, based on multi-scale structural fusion and boundary refinement modeling, introducing a pollution probability mapping generation and consistency optimization mechanism based on a structural prior map. This results in a spatially continuous, boundary-smooth, and structurally reasonable final prediction result. Specifically, the fusion feature map output based on fine-grained boundary structure modeling... Predicting head via convolution Generate a probability value P for each pixel belonging to the sewage region, and introduce a consistency regularization term during the training phase. The predicted probability map is structurally guided to match the candidate region structure, as shown below: , in This represents a consistency regularization term. Indicates the height of the image. Indicates the width of the image. This represents the final structure consistency enhancement map of all pixels. This represents the probability value that all pixels belong to the sewage region. This represents a very small positive number used to prevent the denominator from being zero.
2. The remote sensing wastewater region identification method based on graph structure and multi-stage enhancement according to claim 1, characterized in that, The data preprocessing and characterization enhancement of the acquired remote sensing images includes introducing a spectral consistency normalization strategy. This involves performing spectral domain standardization on the original images to eliminate spectral drift caused by non-target factors. Specifically, to ensure comparability between bands at the same scale, the spectral vectors of all pixels are first normalized using Z-score standardization, converting the pixel values of each band to a form with a mean of 0 and a standard deviation of 1. Simultaneously, to enhance the difference expression and feature discrimination ability between bands, a learnable frequency domain transform enhancement mechanism is introduced. The spectral vector is treated as a one-dimensional signal, and a Fourier transform is performed to extract frequency domain features. Furthermore, to suppress frequency domain interference and enhance texture expression, a Fourier transform-based frequency domain enhancement mechanism is introduced, achieving the preservation of high-frequency structures and the suppression of low-frequency redundancy.
3. The remote sensing wastewater region identification method based on graph structure and multi-stage enhancement according to claim 2, characterized in that, The construction of the anomaly enhancement mapping based on local statistical distribution includes preprocessing and enhancing the remote sensing image tensor. ,in , These are the height and width of the image, respectively. For each band, an anomaly enhancement mapping method based on local statistical distribution deviation is constructed; for each pixel position in the image... Extract its in Centered on, with side length as Local window Calculate the mean vector of all pixels within the window in the spectral dimension. Covariance Matrix : , , in Indicates the position within a local window spectral vector, This indicates the number of pixels in a local window. This represents the mean vector of all pixels within the window along the spectral dimension. Let represent the covariance matrix.
4. The remote sensing wastewater region identification method based on graph structure and multi-stage enhancement according to claim 3, characterized in that, The mechanism for enhancing stable regional response based on the structural consistency enhancement of contaminated areas includes structural candidate maps. As an initial saliency cue, and also based on the preprocessed image tensor A structural feature representation is constructed for cross-temporal relative alignment and consistency scoring. To measure the structural consistency between the current frame and the reference frame in the candidate region, the cosine similarity of the vector angle is introduced as a consistency metric. Subsequently, to avoid noise propagation caused by directly using the original structural graph, only the structural consistency in the candidate region is statistically analyzed, and the final structural consistency enhancement mask is defined.
5. The remote sensing wastewater region identification method based on graph structure and multi-stage enhancement according to claim 4, characterized in that, The graph attention network modeling and structure-enhanced reasoning include introducing a graph attention network and dynamically adjusting the information transmission intensity between adjacent nodes through a multi-head self-attention mechanism, thereby achieving cross-regional feature-enhanced reasoning. The constructed node set is... Each node has the following initial features: The adjacency weight matrix is , No. Attention head to node The update is represented as: , in Indicates the first Node in the head For nodes Attention weights The characteristic linear transformation matrix, For attention parameters, This represents vector concatenation. For nodes The set of adjacent nodes, LeakyReLU is a non-linear activation function used to enhance the model's ability to express negative weights. Represents a node initial characteristics, This represents the initial characteristics of each node. Represents a node The initial characteristics of the neighboring nodes.
6. The remote sensing wastewater region identification method based on graph structure and multi-stage enhancement according to claim 5, characterized in that, The spatial consistency enhancement and segmentation result optimization include introducing a graph enhancement mapping mechanism and a structure preservation regularization term to improve the spatial consistency and representational integrity of the final segmentation graph. The node features output by the graph attention network are... , indicating the first after structural reasoning The semantic representation of each contaminated sub-block is obtained, and then each node feature is assigned to its corresponding pixel set through a region inverse mapping operation. The enhanced feature map is obtained. Subsequently, the enhanced feature map and fused features will be... Perform splicing and fusion, and input into a lightweight decoder layer. To generate a probability map of the final contaminated area To mitigate the potential feature shift risk during graph reasoning, a structure preservation regularization term is introduced to measure the final predicted graph. Note the input probability graph compared to the original graph. Spatial structural consistency between them is defined by the structural consistency regularization term as a weighted KL divergence form: , in, This represents a regularization term for spatial structure consistency. This represents the probability value of a polluted area. A predicted probability map representing the original contaminated area. This represents a predicted probability map of the final contaminated area. This represents the edge saliency map, defined in the previous module, used to strengthen the constraint on the consistency of boundary regions.
7. A remote sensing wastewater region identification system based on graph structure and multi-stage enhancement, executing the remote sensing wastewater region identification method based on graph structure and multi-stage enhancement as described in claim 1, characterized in that, include: The data acquisition module is configured to acquire remote sensing images; The preprocessing module is configured to perform data preprocessing and characterization enhancement on the acquired remote sensing images; The initial screening module is configured to perform initial screening of salient wastewater areas in the enhanced remote sensing image, including: constructing anomaly enhancement maps based on local statistical distribution; extracting pollution candidate maps driven by spatial structure priors; and enhancing the response of stable areas based on the pollution area structure consistency enhancement mechanism. The identification module is configured for high-precision segmentation and identification of wastewater areas, including: construction of multi-resolution residual pyramid structure; fine-grained boundary structure modeling and uncertainty suppression; generation of wastewater distribution probability map and optimization of structural consistency. The modeling module is configured to perform spatial relationship modeling of wastewater regions based on graph attention, including graph structure construction and node relationship initialization; graph attention network modeling and structure enhancement inference; spatial consistency enhancement and segmentation result optimization. The output module is configured to output the results.