Remote sensing water quality pollution detection method based on super-resolution mixed GAN

Abstract

The invention relates to the technical field of water quality detection, and further relates to a remote sensing water quality pollution detection method based on a super-resolution hybrid GAN, and the method comprises the steps: 1, inputting a remote sensing reflectivity image into a super-resolution hybrid generative adversarial network, and carrying out the collaborative reconstruction through a generator and a discriminator, so as to obtain a high-resolution remote sensing reflectivity image; step 2, carrying out weighted summation on the high-resolution remote sensing reflectivity image according to wave bands by utilizing the wave band pollutant absorption sensitivity coefficient calibrated by an experiment to obtain an original concentration estimation value of each pollutant; 3, obtaining the final concentration of each pollutant based on the confrontation consistency loss of the super-resolution hybrid generative adversarial network in combination with the maximum absorption peak wavelength of the pollutant; and 4, generating a comprehensive water quality pollution index. According to the method, the targets of high precision, high reliability and high adaptability of pollution detection under the driving of the remote sensing image are effectively realized.

Description

Technical Field

[0001] The present invention belongs to the field of water quality detection technology, and specifically relates to a remote sensing water pollution detection method based on super-resolution hybrid GAN. Background Art

[0002] As water environment problems become increasingly prominent, the application of remote sensing technology in water pollution monitoring has gradually become a research hotspot in the field of environmental monitoring. Multispectral or hyperspectral images obtained through remote sensing platforms can be used to conduct non-contact observations of large-scale water bodies, and have advantages in temporal and spatial coverage that traditional point sampling cannot match. By utilizing the correlation between the reflectance of each band in remote sensing images and the concentration of pollutants, it is possible to identify pollutant types, estimate concentrations, and analyze spatiotemporal changes. This has been widely verified in various types of water bodies such as lakes, rivers, and reservoirs. However, the current mainstream remote sensing water quality detection methods still have obvious deficiencies in accuracy, resolution, and model robustness, especially in the coordinated identification of multiple pollutants and the restoration of high-resolution spatial details. There is still room for breakthroughs.

[0003] Currently, remote sensing water quality monitoring primarily relies on a combination of multispectral imagery and empirical or semi-empirical inversion models. These methods, typically based on laboratory or field sampling data, perform regression modeling between the reflectance of each band in the remote sensing image and specific water quality parameters (such as chemical oxygen demand, total phosphorus, total nitrogen, and chlorophyll). The most common modeling approaches include multivariate linear regression, principal component regression, partial least squares regression, and support vector machines. While these methods have some practicality, they are highly dependent on the representativeness of the sample data and the consistency of the scene. Under conditions of varying water types, varying lighting conditions, and significant sediment disturbance, the models' generalization ability is poor, and they are prone to overestimation or underestimation of concentrations.

[0004] Furthermore, remote sensing images are subject to the physical limitations of imaging equipment, especially when optical sensors have insufficient spatial resolution, making it difficult to accurately identify the distribution of subtle pollutants in water bodies. Most satellite remote sensing platforms provide low-resolution water quality remote sensing data. For example, the pixel side length of medium-resolution images is generally 10 to 30 meters, making it difficult to achieve effective observation coverage in complex water systems, narrow river channels, or localized sewage outfalls. This results in traditional models often experiencing information loss or oversmoothing when interpolating concentrations, making it difficult to accurately depict the boundaries of pollutant distribution. Summary of the Invention

[0005] The main purpose of this paper is to provide a remote sensing water pollution detection method based on a super-resolution hybrid GAN. This method uses a generator and a discriminator to collaboratively reconstruct high-resolution reflectance information from remote sensing images, combines pollutant absorption sensitivity coefficients for concentration estimation, and introduces adversarial consistency loss for pixel-level correction. Finally, a comprehensive water pollution index is constructed based on ecotoxicity weights. This method effectively improves the spatial accuracy and inversion reliability of remote sensing images, enhances the physical consistency of pollutant identification and the ability to express ecological risks, and achieves the goals of high-precision, high-reliability, and high-adaptability pollution detection driven by remote sensing imagery.

[0006] In order to solve the above problems, the technical solution of the present invention is achieved as follows: A remote sensing water pollution detection method based on super-resolution hybrid GAN, the method comprising: Step 1: Obtain a remote sensing reflectance image of the target area through a remote sensing imaging device; input the remote sensing reflectance image into a super-resolution hybrid generative adversarial network, and obtain a high-resolution remote sensing reflectance image through collaborative reconstruction by the generator and discriminator; Step 2: Using the experimentally calibrated band pollutant absorption sensitivity coefficient, perform weighted summation of the high-resolution remote sensing reflectance image by band to obtain the original estimated concentration value of each pollutant; Step 3: Based on the adversarial consistency loss of the super-resolution hybrid generative adversarial network and the maximum absorption peak wavelength of the pollutant, pixel-level correction is performed on the original estimated concentration of each pollutant to obtain the final concentration of each pollutant; Step 4: According to the ecotoxicity weight and water quality standard limit, the final concentration of each pollutant is weighted and normalized to generate a comprehensive water quality pollution index to represent the overall water quality pollution level.

[0007] Furthermore, step 1 specifically includes: obtaining a low-resolution remote sensing reflectance image, and inputting the low-resolution remote sensing reflectance image into the generator of the trained super-resolution hybrid generative adversarial network; the generator outputs a high-frequency residual signal for each pixel and each band; using the discriminator to calculate the true value probability of the water body for the same pixel, and performing confidence modulation on the high-frequency residual signal; then obtaining the corresponding point spread function scale scaling coefficient based on the square of the ratio of the half-width of the point spread function of each band to the half-width of the highest resolution reference point spread function, and using the point spread function scale scaling coefficient to perform amplitude correction on the modulated high-frequency residual signal; the amplitude-corrected high-frequency residual signal is summed band by band with the original low-resolution remote sensing reflectance image to obtain a super-resolution remote sensing reflectance image.

[0008] Furthermore, in the super-resolution remote sensing reflectance image of step 1, the super-resolution reflectance corresponding to each pixel is: in, Indicates that in high-resolution remote sensing reflectance images, the row index is , the column index is The pixel has a wavelength of Super-resolution reflectivity at ; In the remote sensing reflectance image, the row index is , the column index is The pixel has a wavelength of Reflectivity at ; represents the high-frequency reflectivity residual predicted by the generator; The true value probability of water body calculated by the discriminator; represents the PSF scaling factor of the remote sensing imaging device; The center wavelength of the band is The point spread function of the band when ; For reference, the highest resolution.

[0009] Furthermore, step 2 specifically includes: for each pollutant, using laboratory standard water sample spectrum calibration to obtain the corresponding multi-band reflectance absorption sensitivity coefficient, multiplying the multi-band reflectance absorption sensitivity coefficient with the super-resolution remote sensing reflectance image item by item in the same pixel and the same band, and summing them band by band in the entire band range to obtain the original concentration estimate of each pollutant.

[0010] Furthermore, step 3 specifically includes: calculating the adversarial consistency loss of the super-resolution hybrid generative adversarial network within the same pixel scale, multiplying the adversarial consistency loss by the ratio of the maximum spectral absorption peak wavelength of the pollutant and the fixed reference wavelength to obtain a correction denominator; dividing the original concentration estimate by the correction denominator to obtain the final concentration of the pollutant after adversarial consistency correction.

[0011] Furthermore, the method for calculating the adversarial consistency loss of the super-resolution hybrid generative adversarial network at the same pixel scale specifically includes: for the target pixel and its eight neighborhoods, respectively extracting the generated branch input output by the generator and the real branch input extracted from the high-resolution remote sensing reflectance image, and aligning the two one by one in terms of the number of bands and pixel positions; performing convolution degradation and downsampling on each band of the generated branch input according to the pre-calibrated corresponding band point spread function, and then realigning it with the real branch input through bicubic interpolation to obtain a physical consistency control block; synchronously sending the real branch input and the physical consistency control block into the convolutional attention network of the same discriminator, retaining only the two-dimensional probability map output by the global average pooling layer, and indexing the real probability of the center pixel and the generated probability of the center pixel. rate; calculate the cross-band first-order difference spectral gradient of the true branch input and the physical consistency control block respectively, extract the difference in the gradient modulus of the central pixel and combine it with the gradient direction cosine to obtain the multi-band gradient residual scalar; calculate the second-order Laplace spatial gradient of the generated branch input and the true branch input, extract the difference in the Laplace value of the central pixel and compare it with the neighborhood to obtain the Laplace contrast residual, and normalize it according to the spectral reflectance variance; determine the band confidence weight, perform weighted averaging on the spectral gradient residual and the spatial coherence penalty, and jointly form a composite loss molecule with the true generation probability difference of the central pixel; calculate the local water texture scale factor based on the difference in the visible light near-infrared water body index of the target pixel and its neighborhood, perform linear normalization on the composite loss molecule, and output the adversarial consistency loss.

[0012] Furthermore, from the high-resolution remote sensing reflectance image, the real spectral space block is extracted for the target pixel and its eight neighborhoods as the real branch input; at the same central pixel position, the inferred spectral space block of the corresponding size is synchronously extracted from the high-resolution residual signal output by the self-generator as the generation branch input; the center pixel is explicitly indexed on the two-dimensional probability map to obtain the true probability of the center pixel and the generation probability of the center pixel; the Laplace value difference of the center pixel is extracted and compared with the four-connected center of the neighborhood to obtain the Laplace contrast residual; the band confidence weight is determined based on the inverse of the radiation signal-to-noise ratio of each band.

[0013] Furthermore, for row index , the column index is The comprehensive water pollution index of the pixel is: Among them, when If it is greater than 1, it means that the pollution exceeds the standard at that pixel; The most toxic of all pollutants , used to normalize the toxicity of each pollutant so that the weight ratios are all within within the scope; Indicates the The concentration of a pollutant that causes 50% mortality of a certain indicator aquatic organism within 96 hours under experimental conditions reflects the strength of its ecotoxicity. The smaller the value, the stronger the toxicity. The row index is , the column index is Adversarial consistency loss of pixels; Indicates the The upper limit of the permissible concentration of the pollutant.

[0014] The present invention's remote sensing water pollution detection method based on a super-resolution hybrid GAN has the following beneficial effects: This approach significantly improves the spatial resolution, pollutant concentration inversion accuracy, and pollution risk expression capabilities of remote sensing images for water quality monitoring. Compared with traditional techniques based on direct inversion of low-resolution images, this method effectively enhances image detail and suppresses noise interference by introducing a super-resolution reconstruction mechanism that collaboratively optimizes the generator and discriminator. This allows for the characterization of areas such as water body boundaries, localized pollution zones, and small-scale discharge outlets with higher spatial accuracy. Furthermore, the present invention utilizes experimentally calibrated band-based pollutant absorption sensitivity coefficients to implement physical modeling of pollutant concentrations, improving the interpretability and cross-regional adaptability of concentration estimates. By introducing an adversarial consistency loss into the discriminator output during concentration correction, the inversion results not only conform to the pollutant reflectance characteristics in the spectral dimension but also closely resemble the actual water body morphology in terms of spatial structure, significantly reducing inversion bias caused by image reconstruction errors. Furthermore, the present invention fully considers the differences in pollutant ecotoxicity in the construction of a comprehensive water pollution index. A normalized weighting mechanism effectively reflects the actual hazard level of pollutants, enabling a quantitative expression of pollution risk. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A schematic diagram of a method flow chart for remote sensing water pollution detection based on super-resolution hybrid GAN provided by an embodiment of the present invention; Figure 2 A graph showing a comparison of super-resolution reconstruction effects provided by an embodiment of the present invention; Figure 3 A graph showing the accuracy of multi-band pollutant concentration detection experiments provided by an embodiment of the present invention; Figure 4 A schematic flowchart of the adversarial consistency loss calculation provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0016] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0017] refer to Figure 1 :A remote sensing water pollution detection method based on super-resolution hybrid GAN, the method comprising: Step 1: Obtain a remote sensing reflectance image of the target area through a remote sensing imaging device; input the remote sensing reflectance image into a super-resolution hybrid generative adversarial network, and obtain a high-resolution remote sensing reflectance image through collaborative reconstruction by the generator and discriminator; Step 2: Using the experimentally calibrated band pollutant absorption sensitivity coefficient, perform weighted summation of the high-resolution remote sensing reflectance image by band to obtain the original estimated concentration value of each pollutant; Step 3: Based on the adversarial consistency loss of the super-resolution hybrid generative adversarial network and the maximum absorption peak wavelength of the pollutant, pixel-level correction is performed on the original estimated concentration of each pollutant to obtain the final concentration of each pollutant; Step 4: According to the ecotoxicity weight and water quality standard limit, the final concentration of each pollutant is weighted and normalized to generate a comprehensive water quality pollution index to represent the overall water quality pollution level.

[0018] The remote sensing water pollution detection method based on super-resolution hybrid GAN disclosed in the present invention is aimed at multi-pollutant concentration inversion and comprehensive risk assessment scenarios. It fully integrates remote sensing image reconstruction technology and adversarial learning framework, breaking through the limitations of traditional spectral-concentration linear mapping that relies on single-resolution images and cannot suppress multi-scale artifacts and sensor noise. In the overall implementation process, a remote sensing imaging device is first used to obtain a remote sensing reflectance image of the target area. The image retains the true surface reflectance information after radiometric calibration, geometric correction and atmospheric correction; then the remote sensing reflectance image is input into a super-resolution hybrid generative adversarial network. The generator explicitly introduces a cross-band feature fusion module and a residual attention mechanism in the receptive domain to capture fine-grained water textures. The discriminator dynamically measures the generated-true alignment error through a multi-scale perception branch. The iterative game between the two can effectively restore high-frequency details degraded by diffraction blur, mixed pixel effects and quantization noise, thereby outputting a high-resolution remote sensing reflectance image. To ensure that the reconstruction results are consistent with the physical characteristics of the sensor, separable convolution kernels of the point spread functions of each band are embedded in the network degradation model, and a regularization term based on spectral-spatial consistency is added to the loss function to reduce the phantom texture and over-sharpening phenomena commonly seen in the super-resolution process.

[0019] After obtaining the high-resolution remote sensing reflectance image, the present invention establishes the correspondence between the spectral response and the water quality index through the experimentally calibrated band pollutant absorption sensitivity coefficient. The absorption sensitivity coefficient comes from the difference in the measured reflectance of water samples of known concentration under laboratory conditions. The spectral fitting technology is used to regress and solve the characteristic absorption bands of different pollutants, which can retain both wavelength dependence and spectral shape dependence. The above-mentioned absorption sensitivity coefficient and the high-resolution remote sensing reflectance image are weighted and summed pixel by pixel according to the band, and the original estimated value of the concentration of each pollutant can be directly obtained. In order to reduce the uncertainty caused by cross-sensor differences and atmospheric residuals, the present invention introduces adaptive weight normalization based on the band signal-to-noise ratio in the summation link, and uses a water-non-water mask to screen out high-reflection interference pixels on the land surface.

[0020] The design of the adversarial consistency loss is the key that distinguishes the present invention from the prior art. This loss term measures the difference between the real image block and the generated image block at the pixel level. First, the real probability and the generated probability output by the discriminator are used to calculate the pixel authenticity deviation. At the same time, an energy consistency index based on band correlation is constructed in the frequency domain to measure the fidelity of the generated result in the spectral shape. Then, the ratio of the maximum absorption peak wavelength of the pollutant to the reference wavelength is comprehensively considered, and the above two types of errors are mapped into penalty coefficients for the original estimated concentration value to achieve pixel-level correction of the concentration of each pollutant. Since the adversarial consistency loss shares the gradient with the network training process, the correction can be completed in the inference stage without additional computational overhead, so that the concentration estimation significantly suppresses the diffusion of artifacts while ensuring physical consistency.

[0021] During the implementation of the method, a remote sensing reflectance image of the target area is first obtained through a remote sensing imaging device. The remote sensing reflectance image is generally multispectral data with low spatial resolution. Since the remote sensing imaging process is limited by physical constraints such as optical system resolution, sensor size, and orbit height, the low-resolution image cannot meet the spatial detail expression required for high-precision inversion of water pollution. Therefore, it is necessary to input the remote sensing reflectance image into a super-resolution hybrid generative adversarial network, relying on the generator and discriminator in the network to work together to perform detail enhancement and structural restoration on the original image. Among them, the generator is responsible for predicting high-frequency compensation information from the low-resolution reflectance image, while the discriminator guides the generator to optimize the reconstruction result by comparing the pixel-level distribution and spatial consistency of the real image and the generated image, thereby obtaining a remote sensing reflectance image that is closer to the real high-resolution image.

[0022] After generating high-resolution remote sensing reflectance imagery, a preliminary estimate of pollutant concentrations is performed. For each pollutant, the reflectance response of laboratory standard water samples is measured in different bands to obtain the pollutant's absorption sensitivity coefficients in each band. These absorption sensitivity coefficients are then weighted and summed with the pixel reflectance values ​​for the corresponding bands in the high-resolution remote sensing reflectance imagery to obtain a raw estimate of the pollutant concentration at each pixel. This process establishes a quantitative mapping from high-resolution remote sensing imagery to pollutant concentration estimates, reflecting the physical connection between remote sensing spectral characteristics and pollutant distribution.

[0023] To further improve the accuracy of pollutant concentration estimation results, this method introduces an adversarial consistency loss based on a super-resolution hybrid generative adversarial network as a correction factor, and combines the maximum absorption peak wavelength of each pollutant to perform pixel-level concentration correction. The adversarial consistency loss is used to measure the deviation between the generator output and the true high-resolution image in spatial and spectral characteristics. By comparing the true probability and the generated probability at the pixel level, it reflects the degree of deviation of the current pixel in terms of realism and discriminant consistency. By combining this adversarial consistency loss with the spectral response characteristics of the pollutant, a spatially sensitive adjustment mechanism for the original concentration estimate is constructed. This process emphasizes the comprehensive balance between the model's image structure preservation and spectral response accuracy, and effectively suppresses local concentration anomalies caused by remote sensing data noise, artifacts, or high-reflection background.

[0024] After obtaining the final concentration values ​​of each pollutant, a comprehensive water quality pollution index is calculated. Considering the varying toxic effects of different pollutants on aquatic organisms within ecosystems, this method incorporates an ecotoxicity weighting mechanism to ensure scientific and risk-based pollution assessment. Specifically, an ecotoxicity normalization factor is constructed for each pollutant based on the concentration at which each pollutant causes a 50% mortality rate for typical aquatic organisms under experimental conditions (i.e., the median lethal concentration). Subsequently, the final concentration of the pollutant is compared with the corresponding national or regional water quality standard limit to determine the degree of exceedance. This is then weighted and aggregated using the ecotoxicity weights to ultimately produce a comprehensive water quality pollution index for each remote sensing pixel. This pollution index is a dimensionless parameter, facilitating regional comparison, grading, and visualization.

[0025] It should be noted that in the implementation of the method of the present invention, the calculation of the adversarial consistency loss must be strictly performed within the same pixel scale, and the influence of neighborhood context information on the credibility of the discrimination must be considered. In addition, the absorption sensitivity coefficient of each pollutant should be established based on the measured water samples, and ensure that it is accurately matched with the band response function of the remote sensing image to ensure the physical validity of the concentration estimation process. The matching ratio between the maximum absorption peak wavelength of the pollutant and the reference wavelength also needs to be adjusted in combination with the spectral configuration of the remote sensing system. In order to ensure that the spectral fidelity and spatial details of the generated image achieve high-quality output, the super-resolution hybrid generative adversarial network should combine multiple loss functions during the training process, including adversarial loss, reconstruction loss, and spectral consistency loss, so as to improve the generalization ability and application stability of the model as a whole.

[0026] Furthermore, step 1 specifically includes: obtaining a low-resolution remote sensing reflectance image, and inputting the low-resolution remote sensing reflectance image into the generator of the trained super-resolution hybrid generative adversarial network; the generator outputs a high-frequency residual signal for each pixel and each band; using the discriminator to calculate the true value probability of the water body for the same pixel, and performing confidence modulation on the high-frequency residual signal; then obtaining the corresponding point spread function scale scaling coefficient based on the square of the ratio of the half-width of the point spread function of each band to the half-width of the highest resolution reference point spread function, and using the point spread function scale scaling coefficient to perform amplitude correction on the modulated high-frequency residual signal; the amplitude-corrected high-frequency residual signal is summed band by band with the original low-resolution remote sensing reflectance image to obtain a super-resolution remote sensing reflectance image.

[0027] First, a low-resolution remote sensing reflectance image covering the target area is obtained from a remote sensing imaging device. The image is a collection of multi-band two-dimensional images, which are usually radiometrically calibrated in reflectance units. Due to the limited spatial resolution of this image, it has obvious deficiencies in expressing the spatial distribution of water pollution. Therefore, the low-resolution remote sensing reflectance image needs to be input into the generator of a pre-trained super-resolution hybrid generative adversarial network. The generator adopts a deep convolutional network structure, which can learn the spatial distribution between pixels and the spectral correlation between bands from the original image, and output a high-frequency residual signal for each pixel and its corresponding band. This high-frequency residual signal represents the result of the generator's reconstruction of the detail information lost in the original image due to point spread function blurring, including important features such as texture edges, local contrast, and reflectance jumps between bands.

[0028] Subsequently, the high-frequency residual signal output by the generator is fused with the calculation results of the discriminator. The discriminator is the discriminative subnetwork in the super-resolution hybrid generative adversarial network, which has the ability to identify local differences between the real remote sensing image and the generated image. For each pixel, a recognition model based on spatial neighborhood and spectral characteristics is established within the discriminator to output the true confidence of the water pixel, that is, the water body true value probability. The water body true value probability is used to perform confidence modulation on the high-frequency residual signal output by the generator. Specifically, it multiplies the original residual signal with the probability result output by the discriminator to enhance the features of the credible area and suppress the features of the uncredible area, thereby improving the physical consistency and content stability of the reconstructed image.

[0029] During the actual amplitude recovery process, the modulated high-frequency residual signal must be scaled and normalized using the point spread function (PSF) of the remote sensing imaging system. Since remote sensing imaging systems exhibit varying optical response characteristics across different bands, and the spatial resolution of each band varies due to factors such as lens diffraction, filter thickness, and sensor structure, it is necessary to normalize the PSF at half maximum (FWHM) of each band against the FWHM of the system's reference highest-resolution PSF. The square of the ratio of these two factors is used as a scaling factor to construct the PSF scaling coefficient. This coefficient is used to correct the amplitude of the high-frequency residual signal, ensuring that the geometric scale and radiation amplitude of the resulting high-resolution information are consistent with the actual observation conditions of each band, thus preventing inconsistent reconstructed information or spatial resolution mismatches between different bands.

[0030] After amplitude correction is completed, the corrected high-frequency residual signal is summed band by band with the original low-resolution remote sensing reflectance image. This summing process represents the restoration and enhancement of the details of the original image content, ultimately generating a set of super-resolution remote sensing reflectance images with higher spatial resolution, richer texture structure, and stronger inter-band consistency, providing a more reliable data basis for the subsequent fine identification and concentration inversion of water quality pollutants. In the above implementation process, special attention should be paid to the stability of the parameter update mechanism between the generator and the discriminator to avoid mode collapse or gradient oscillation during training. At the same time, the true value probability of the water body output by the discriminator needs to be regularized to ensure that its range is limited to a reasonable range. The measurement of the half-height width of the point spread function should be combined with the specific model of the remote sensing imaging device and the experimental calibration results to ensure that the point spread function scaling coefficient is physically real and applicable.

[0031] Furthermore, in the super-resolution remote sensing reflectance image of step 1, the super-resolution reflectance corresponding to each pixel is: in, Indicates that in high-resolution remote sensing reflectance images, the row index is , the column index is The pixel has a wavelength of Super-resolution reflectivity at ; In the remote sensing reflectance image, the row index is , the column index is The pixel has a central wavelength of Reflectivity at ; represents the high-frequency reflectivity residual predicted by the generator; The true value probability of water body calculated by the discriminator; represents the PSF scaling factor of the remote sensing imaging device; The center wavelength of the band is The point spread function of the band when ; For reference, the highest resolution.

[0032] This relationship explicitly converts the original low-resolution remote sensing reflectance image The residual of the high-frequency reflectivity output by the generator Superposition is performed and the water body true value probability from the discriminator is introduced as the confidence modulation factor, and through This point diffusion function scaling factor achieves amplitude correction for band resolution differences. Its working mechanism is that the generator first predicts the local spatial context of the low-resolution pixel and the cross-band spectral pattern. , the residual numerically represents the detail amplitude lost by optical diffraction blur; then the discriminator gives the same pixel , which is between zero and one after Sigmoid mapping, can be regarded as the credibility weight of the pixel in the real remote sensing statistical distribution. Therefore, the denominator in the formula is It constitutes an adaptive suppression mechanism for high-frequency residual signals, so that the generator can fully retain details in high-confidence areas and automatically converge to zero in low-confidence areas to avoid artifact diffusion.

[0033] On the other hand, different bands are affected by the differences in lens diffraction limit, filter refractive index and detector spacing. To ensure that the detail amplitudes restored in each band have a uniform physical scale, the present invention measures the half-height width of the point spread function of the reference highest resolution band. Full width at half maximum of the target band The square ratio of This factor plays a role in scaling the amplitude of the high-frequency residual signal, which not only makes up for the loss of details caused by diffraction in the long-wave band, but also suppresses the excessive sharpening that may occur in the short-wave band, thereby ensuring In the actual deployment process, the generator and the discriminator need to be updated alternately on the same training image set. In addition to the adversarial loss, the generator loss function needs to be superimposed with spectral consistency constraints, and the discriminator needs to use the joint distribution of real images and synthetic images for discrimination training to ensure Reflects the pixel authenticity probability on a global scale.

[0034] The point spread function half-width parameters need to be obtained through laboratory precision optical testing or on-orbit star observation and stored in a table during the image preprocessing stage to ensure The calculation is accurate and reliable. The entire super-resolution reconstruction process uses pixels as processing units, keeping the spatial topology unchanged and avoiding the errors introduced by coordinate interpolation; at the same time, each band is applied independently during the generation process. Scaling coefficients are used to ensure that the radiometric calibration relationship between spectral channels is not destroyed. It should be noted that in highly reflective bright spots or shadow areas, the true value probability of water output by the discriminator may appear oversaturated or undersaturated. In the training phase, the present invention adjusts the slope of the Sigmoid curve by adding a temperature coefficient to make The distribution of is smoother to prevent the local gradient explosion from causing convergence instability. In addition, in order to avoid the spectrum jump caused by the large difference in point spread functions of different bands, the present invention adds a bandpass filter kernel at the output end of the high-frequency residual signal to limit the highest available frequency of the predicted content so that it does not exceed The upper limit of the resolution that can be supported. Through the above means, the final output In terms of spatial details, spectral morphology and radiation amplitude, it is consistent with the actual capabilities of remote sensing imaging hardware, providing a high-quality data basis for subsequent pollutant concentration estimation, and significantly improving the stability, accuracy and physical credibility of the entire remote sensing water pollution detection process.

[0035] Furthermore, step 2 specifically includes: for each pollutant, using laboratory standard water sample spectrum calibration to obtain the corresponding multi-band reflectance absorption sensitivity coefficient, multiplying the multi-band reflectance absorption sensitivity coefficient with the super-resolution remote sensing reflectance image item by item in the same pixel and the same band, and summing them band by band in the entire band range to obtain the original concentration estimate of each pollutant.

[0036] First, it is necessary to carry out spectral measurements of laboratory standard water samples for each pollutant. In specific implementation, by collecting standard water samples with a certain concentration gradient, the reflectivity change trend at the central wavelength of multiple remote sensing bands is recorded in the spectral measurement instrument, and combined with the known water sample concentration, linear fitting, principal component regression or partial least squares method are used to determine the absorption sensitivity coefficient between the pollutant concentration and the reflectivity of each band. The absorption sensitivity coefficient reflects the proportional relationship between the spectral absorption intensity and concentration change of a specific pollutant in each band. It has a clear mass concentration unit, usually in milligrams per liter. The absorption sensitivity coefficient is the result of empirical modeling of pollutants under specific spectral response conditions. It has band dependence and pollutant type dependence. Different pollutants have obvious differences in absorption characteristics in different bands. Therefore, each pollutant needs to be experimentally calibrated separately to establish a set of absorption sensitivity coefficient matrices corresponding to bands and pollutants.

[0037] The above absorption sensitivity coefficient matrix is ​​fused with the super-resolution remote sensing reflectance image generated in step 1 to ensure a one-to-one correspondence between the data in the pixel space position and the band spectral position. That is, for each pixel position in the image and the reflectance values ​​of each band it contains, the absorption sensitivity coefficient of the corresponding pollutant is extracted item by item and multiplied with the reflectance value of the band item by item. This process is a pixel-by-pixel and band-by-band calculation operation, and the intermediate results obtained represent the quantitative results of the spectral response of the pollutant to the concentration in each band. Subsequently, the response results of all bands are summed band by band to accumulate the original estimated concentration value of the target pollutant at the pixel.

[0038] In actual operation, in order to ensure the matching consistency between the multi-band reflectance absorption sensitivity coefficient and the super-resolution remote sensing reflectance image, it is necessary to ensure that the spectral band center wavelength, bandwidth and sampling resolution of the remote sensing image are consistent with the spectral device parameters used in the experimental calibration or maintain consistency after interpolation and normalization. In addition, the super-resolution remote sensing reflectance image should have completed atmospheric correction and radiation calibration processing, and the reflectance values ​​of all bands should be the apparent reflectance or surface reflectance of the ground object to avoid interference from external factors such as atmospheric path radiation and changes in solar altitude angle. For the extraction of absorption sensitivity coefficients, a sample system should be established under multiple representative water scenes, and the influence of variables such as water color type, water transparency, and suspended matter content on the spectral response should be considered to avoid systematic deviations in concentration estimation results under different environmental conditions due to insufficient model generalization.

[0039] It is worth noting that the multi-band weighted summation concentration estimation model is essentially a linear spectral mixing inversion process. Therefore, when the spectral response curve of the target pollutant exhibits nonlinear characteristics or cross-interference effects, the estimation accuracy should be improved by adjusting the band selection, introducing nonlinear transformations, or utilizing auxiliary parameters. In addition, when some bands are saturated or have low signal-to-noise ratios, robustness should be improved by setting band masking or dynamically adjusting band weighting strategies. All raw concentration estimates maintain mass concentration units consistent with the absorption sensitivity coefficient and are retained in subsequent steps as the basis for the final pollutant concentration estimate.

[0040] Furthermore, step 3 specifically includes: calculating the adversarial consistency loss of the super-resolution hybrid generative adversarial network within the same pixel scale, multiplying the adversarial consistency loss by the ratio of the maximum spectral absorption peak wavelength of the pollutant and the fixed reference wavelength to obtain a correction denominator; dividing the original concentration estimate by the correction denominator to obtain the final concentration of the pollutant after adversarial consistency correction.

[0041] First, using the pixel as the processing unit, the original concentration estimate of the pixel in the super-resolution remote sensing reflectance image is extracted. The true water value probability output by the discriminator at the pixel location and the generated image consistency evaluation result are called to construct the adversarial consistency loss for the pixel. This loss value can be regarded as an indicator of the deviation between the generated image at that location and the real remote sensing image in terms of spatial structure and spectral morphology. The larger the value, the more likely it is that there will be problems such as detail deviation, insufficient discriminant confidence, or spatial artifact superposition in the super-resolution restoration process at that location. Therefore, this loss term is used as a suppression factor in concentration inversion to adjust the original concentration estimate results in the uncertain area and prevent abnormal amplification of pollutant concentration.

[0042] After the calculation of the anti-consistency loss is completed, in order to further enhance its physical rationality in the spectral dimension, the present invention multiplies the loss term with the ratio between the maximum absorption peak wavelength of the pollutant in the remote sensing band and the fixed reference wavelength to form a correction denominator with spectral dependence characteristics. The maximum absorption peak wavelength refers to the central wavelength where the pollutant has the strongest absorption in the actual water spectrum, which is usually obtained by laboratory spectral measurement and represents the sensitivity of the pollutant to the spectral information of a specific band. The fixed reference wavelength is usually selected as the center value of the green or near-infrared band in the remote sensing image as a band normalization reference for constructing a unified scale factor. The anti-consistency loss is modulated by the ratio to perform band response modulation, and different pollutants obtain different correction factors according to their spectral activity, thereby enhancing the pollutant specificity and spectral dependence of the correction process.

[0043] The original concentration estimate is then divided by the constructed correction denominator to obtain the final concentration value for the corresponding pollutant at that pixel. This concentration value is corrected after applying a spatial consistency penalty and adjusting for spectral sensitivity, and its numerical value better reflects the physical laws governing the distribution of pollutants in real water bodies. This correction mechanism effectively avoids concentration anomalies caused by distortion of high-frequency details, local structural offsets, or overfitting of the inversion model, thereby improving the spatial smoothness and physical interpretability of the concentration estimate.

[0044] During implementation, the adversarial consistency loss should be calculated by fusing the differences in image spatial features and spectral structure gradients to ensure that it has both spatial and spectral discrimination capabilities. The selection of the maximum absorption peak wavelength must be based on the spectral characteristics of the pollutants and the distribution range of the remote sensing band to ensure that the ratio falls within a reasonable dynamic range, thereby avoiding errors in excessive amplification or compression of the final concentration. In addition, to avoid the discriminator outputting an abnormally high probability of misjudgment in edge areas or highly reflective areas, background masks and water body prior enhancement mechanisms can be introduced during the training phase to optimize the discriminator's consistent response performance in real water areas.

[0045] By executing this step, the present invention achieves pixel-level pollutant concentration correction, constructing a high-precision concentration estimation mechanism for super-resolution remote sensing data. This significantly improves the reliability and distribution continuity of pollutant concentration inversion, providing more reliable basic data for subsequent pollution index evaluation. While maintaining the physical structure of the original concentration, the entire process incorporates the discriminative power of a generative adversarial network, achieving a physically consistent closed loop from image data to pollution concentration results in remote sensing inversion.

[0046] Furthermore, the method for calculating the adversarial consistency loss of the super-resolution hybrid generative adversarial network at the same pixel scale specifically includes: for the target pixel and its eight neighborhoods, respectively extracting the generated branch input output by the generator and the real branch input extracted from the high-resolution remote sensing reflectance image, and aligning the two one by one in terms of the number of bands and pixel positions; performing convolution degradation and downsampling on each band of the generated branch input according to the pre-calibrated corresponding band point spread function, and then realigning it with the real branch input through bicubic interpolation to obtain a physical consistency control block; synchronously sending the real branch input and the physical consistency control block into the convolutional attention network of the same discriminator, retaining only the two-dimensional probability map output by the global average pooling layer, and indexing the real probability of the center pixel and the generated probability of the center pixel. rate; calculate the cross-band first-order difference spectral gradient of the true branch input and the physical consistency control block respectively, extract the difference in the gradient modulus of the central pixel and combine it with the gradient direction cosine to obtain the multi-band gradient residual scalar; calculate the second-order Laplace spatial gradient of the generated branch input and the true branch input, extract the difference in the Laplace value of the central pixel and compare it with the neighborhood to obtain the Laplace contrast residual, and normalize it according to the spectral reflectance variance; determine the band confidence weight, perform weighted averaging on the spectral gradient residual and the spatial coherence penalty, and jointly form a composite loss molecule with the true generation probability difference of the central pixel; calculate the local water texture scale factor based on the difference in the visible light near-infrared water body index of the target pixel and its neighborhood, perform linear normalization on the composite loss molecule, and output the adversarial consistency loss.

[0047] In the specific implementation process, the target pixel to be analyzed is first taken as the center, and a two-dimensional neighborhood window consisting of three rows and three columns, totaling nine pixels, is extracted from the high-resolution remote sensing reflectance image. This neighborhood window is simultaneously aligned in the generator output image and the reference real image to serve as the input of the generated branch and the real branch, respectively. To ensure the effectiveness of the comparison and the alignment of the spatial structure, the generated branch and the real branch must have a complete one-to-one correspondence in terms of the number of bands and pixel positions. That is, all band channels must be consistent and spatially sliced ​​with the same row and column indices. The above inputs constitute the spatial-spectral joint pixel block structure required to generate the discriminant path and are the input basis for the entire loss calculation.

[0048] In order to simulate the point spread phenomenon in different bands of actual remote sensing imaging systems and further construct a physical degradation control for the generated image, after obtaining the input of the generation branch, a two-dimensional convolution operation is performed on each band based on the pre-calibrated point spread function. The width of the convolution kernel is based on the wavelength corresponding to the center wavelength of the band. The convolution result simulates the spatial blurring process during band imaging. The convolved image is then spatially downsampled to match the original sensor resolution, forming a degraded image. To ensure that it is fully consistent in size with the true branch input, the downsampled image is reconstructed using bicubic interpolation to restore it to the original high-resolution pixel scale, ultimately resulting in a physical consistency comparison block. This physical consistency comparison block represents the restored image after the generated image has been subjected to a real remote sensing imaging degradation model. Its structure is used to compare various feature differences with the true branch image.

[0049] The real-branch input and the physical consistency check block are simultaneously fed into the weight-shared discriminator convolutional attention network. Standard convolution operations and multi-head attention aggregation are performed, and at the end, a two-dimensional probability map output by the global average pooling layer is retained to quantify the confidence probability distribution that the input region is a real image. On this probability map, the center pixel's index is used to extract the center pixel's true probability and the center pixel's generated probability, obtaining the discriminant difference of the current pixel under the discriminator's semantics. This probability difference serves as one of the loss components of the main discriminant path, indicating the degree of difference between the generated result and the real image at the semantic level.

[0050] To enhance the ability to respond to spectral structural information, the first-order difference spectral gradients of the true branch input and the physical consistency comparison block are calculated in all band dimensions, and the gradient modulus of the central pixel is calculated. Furthermore, the direction cosines between the two gradient vectors are extracted, and the weighted residual of the spectral direction difference is calculated based on the modulus difference, forming a multi-band gradient residual scalar. This residual reflects the degree of detail distortion in the generated image's spectral morphology and plays an important role in discriminating the spectral unmixing of water pollutants.

[0051] To further consider spatial structural fidelity, a second-order Laplacian convolution is performed on both the true and generated branch inputs to obtain a spatial edge response map. The difference in the Laplacian value of the central pixel is then extracted. This difference, combined with the Laplacian responses of the surrounding pixels, forms a local contrast difference indicator, known as the Laplacian contrast residual, reflecting the ability of the generated image to preserve structural details. Taking into account the differences in brightness levels and signal-to-noise ratios across different bands in remote sensing images, this residual is normalized by the spectral reflectance variance to ensure uniform discrimination across bands and suppress error amplification in highly reflective or saturated bands.

[0052] After constructing all discriminant residual terms, band confidence weights are further introduced to improve the response stability of the residual fusion process. These band confidence weights are defined based on the inverse of the radiometric signal-to-noise ratio of each band. Bands with low signal-to-noise ratios receive lower weights to reduce the impact of low-quality bands on the final loss value. The spectral gradient residual and the spatial coherence penalty are weighted averaged using this confidence weight and combined with the true generation probability difference of the central pixel to form a composite loss numerator.

[0053] Finally, to accommodate the complexity of different water textures, the present invention constructs a local water texture scale factor based on the differences in the visible-near-infrared water index between the target pixel and its eight neighbors. This factor reflects the water texture complexity or water color variation in that region. This factor is then used to linearly normalize the aforementioned composite loss numerator, suppressing residual anomalies introduced by complex water regions. The resulting normalization is the final adversarial consistency loss, which is used in subsequent pixel-level correction of pollutant concentrations.

[0054] Furthermore, from the high-resolution remote sensing reflectance image, the real spectral space block is extracted for the target pixel and its eight neighborhoods as the real branch input; at the same central pixel position, the inferred spectral space block of the corresponding size is synchronously extracted from the high-resolution residual signal output by the self-generator as the generation branch input; the center pixel is explicitly indexed on the two-dimensional probability map to obtain the true probability of the center pixel and the generation probability of the center pixel; the Laplace value difference of the center pixel is extracted and compared with the four-connected center of the neighborhood to obtain the Laplace contrast residual; the band confidence weight is determined based on the inverse of the radiation signal-to-noise ratio of each band.

[0055] First, a true spectral space block is extracted from the high-resolution remote sensing reflectance image for the target pixel and its eight-neighborhood region. The spatial dimensions of the spectral space block are three rows and three columns, containing nine pixels, and the spectral dimensions correspond to the total number of bands contained in the remote sensing image. This true spectral space block serves as the input to the true branch, preserving the local spectral structure and spatial distribution characteristics of the target area, and providing the original input for the subsequent true image discrimination path. At the same time, at the same central pixel position, that is, at a position strictly aligned with the spatial index of the true spectral space block, an inferred spectral space block with the same size as the true branch is extracted from the high-resolution residual signal output by the generator as the input to the generative branch. This generative branch input preserves the local spatial texture and band structure information predicted by the generator during the reconstruction process, and together with the true branch, it forms the input basis of the discriminator's dual-branch path.

[0056] The above two branch inputs are synchronously fed into the convolutional attention structure of the super-resolution hybrid generative adversarial network discriminator. After multi-layer spatial convolution, attention feature aggregation and feature mapping processing, the global average pooling layer finally outputs a two-dimensional probability map. Each pixel value of the probability map represents the probability that the corresponding pixel is judged to be a real remote sensing pixel under the discriminator recognition semantics. In order to obtain quantitative indicators for subsequent residual construction, the present invention performs an explicit index operation on the center pixel on the two-dimensional probability map, and directly extracts the true probability and generated probability corresponding to the center pixel under the true branch path and the generated branch path, which are respectively recorded as the center pixel true probability and the center pixel generated probability. The difference between the two probability values ​​is used to characterize the degree of deviation from the statistical authenticity of the generated image at the pixel, and is one of the inputs of the core discriminant path in the adversarial consistency loss.

[0057] In order to further enhance the sensitivity to spatial structural changes, the present invention introduces spatial high-order gradient response features within the central pixel and its neighborhood, that is, the second-order Laplace spatial gradient map is calculated for the real branch input and the generated branch input respectively. The Laplace spatial gradient map is used to capture the edge response and high-frequency structural changes in the image, and is an important basis for judging the consistency of details. After extracting the Laplace response value of the central pixel from the map, it is further compared and analyzed with the Laplace value of the four-connected central pixels in the neighborhood to construct a local contrast residual, which is called the Laplace contrast residual. It quantifies the difference in the degree of preservation of the edge structure of the generated image around the pixel. If the residual value is larger, it means that the degree of restoration of the edge details of the generated image in the pixel is worse, and the weight constraint needs to be increased in subsequent corrections.

[0058] In order to ensure that the various residuals mentioned above have a unified weight system when fused in the spectral dimension, while suppressing the interference of bands with low signal-to-noise ratios on the residual components, the present invention calculates the band confidence weight based on the radiation signal-to-noise ratio of each band. Specifically, the signal-to-noise ratio of each band on multiple training samples is calculated, and its inverse is taken as the weight component. The band with a higher signal-to-noise ratio is assigned a larger confidence weight. This band confidence weight is used in the subsequent weighted fusion operation of multi-band gradient residuals, spatial contrast residuals, and central pixel probability differences to ensure that the final output of the adversarial consistency loss has reasonable sensitivity and stability at the multi-band and multi-structure levels.

[0059] Furthermore, for row index , the column index is The comprehensive water pollution index of the pixel is: Among them, when If it is greater than 1, it means that the pollution exceeds the standard at that pixel; The most toxic of all pollutants , used to normalize the toxicity of each pollutant so that the weight ratios are all within within the scope; Indicates the The concentration of a pollutant that causes 50% mortality of a certain indicator aquatic organism within 96 hours under experimental conditions reflects the strength of its ecotoxicity. The smaller the value, the stronger the toxicity. The row index is , the column index is Adversarial consistency loss of pixels; Indicates the The upper limit of the permissible concentration of the pollutant.

[0060] During the implementation process, after the pollutant concentration correction step is completed, the concentration matrix is ​​constructed according to the pollutant dictionary index order, and all concentration values ​​are uniformly expressed in milligrams per liter as the mass concentration unit. Then, the ecotoxicity database is searched to obtain a set of 96-hour median lethal concentration data consistent with the types of pollutants monitored by remote sensing, and the minimum value is recorded as , and read the corresponding In order to avoid the distortion of the calculation process caused by the difference in numerical scale, it is necessary to Perform missing value verification. If there is no public data on pollutants, reference values ​​of similar chemical structures or similar ecotoxicity parameters can be used, and the source can be recorded in the explanation document. Then, according to the applicable national or regional water environment standards, read the upper limit of the allowable concentration of each pollutant. During the data preparation phase, the concentration units should be consistent and all and Both are milligrams per liter.

[0061] The calculation phase is carried out in a pixel traversal manner: for each pixel position , call the saved Compared with standard limit The excess ratio is composed of the toxicity weight ratio Multiply to complete the calculation of the risk score of a single pollutant, and finally sum up the risk scores of all pollutants to obtain It is recommended to use single-precision floating-point operations in the implementation to reduce the storage pressure caused by large-scale pixel cycles; if deployed on a low-power platform, the memory usage can be reduced by block calculation and cache reuse strategies. To improve the interpretability of Classified as safe. Classified as alert level, It is classified as exceeding the standard level, and different levels are mapped to quantitative color bands in the image rendering process to achieve visual output.

[0062] In terms of precision control, attention should be paid to If the difference in ecotoxicity is too large, the weight of a pollutant may be extremely amplified, causing the index value to be unstable. An upper threshold can be set for the comparison value, for example, limiting it to within ten, and using a percentile truncation strategy for the excess value to maintain the robustness of the index result. In addition, The acquisition of depends on the discriminant consistency output of the super-resolution hybrid generative adversarial network. Its numerical distribution may be affected by image noise, cloud shadows or high light reflections. It is recommended to perform cloud mask and water mask processing on the relevant pixels before concentration correction to avoid non-target areas interfering with the index accuracy.

[0063] In software implementation, The calculation module can be encapsulated as a matrix operation interface, and all pixels can be processed at once through a vectorized programming environment to improve operating efficiency; at the hardware implementation level, it can be combined with a graphics processing unit or a tensor processing unit for parallel computing to meet the needs of generating real-time pollution indexes for large-scale remote sensing images. The raster layer can be used as a base map for visualizing pollution distribution and can also be input into downstream water environment models or river and lake health assessment systems to provide a basis for water quality improvement measures.

[0064] Figure 2 This figure demonstrates a comparison of the reconstruction performance of remote sensing reflectance imagery using the super-resolution hybrid generative adversarial network (GAN) proposed in this invention. The horizontal axis represents wavelength, in nanometers, covering the visible to near-infrared spectral range from 400 to 900 nanometers, encompassing the primary spectral absorption characteristic bands of water pollutants. The vertical axis represents reflectance, ranging from 0.0 to 1.0, representing the spectral reflectance intensity of the water surface at different wavelengths. The figure includes two comparison curves: the dashed line represents the spectral response characteristics of the original low-resolution remote sensing reflectance image, and the solid line represents the high-resolution remote sensing reflectance image processed by the super-resolution hybrid GAN. The comparison of the curves clearly shows that the reflectance variation of the low-resolution remote sensing reflectance image is relatively flat across all bands, with a significant loss of spectral detail. In particular, in the 500-700 nanometer band, the curve exhibits a relatively smooth trend, making it difficult to capture the fine spectral absorption characteristics of water pollutants. In contrast, the spectral curve of the high-resolution remote sensing reflectance image displays richer details, with more distinct absorption and reflection peaks in pollutant-sensitive bands. In particular, the high-resolution curves exhibit sharper spectral response variations in the 400-500 nm and 600-800 nm bands. These detailed features are crucial for the subsequent quantitative inversion of pollutant concentrations. Through the collaborative reconstruction of the generator and discriminator, the high-resolution remote sensing reflectance imagery effectively preserves the high-frequency information of the water spectrum, providing a reliable data foundation for the accurate detection of water pollutant concentrations.

[0065] Figure 3 This figure demonstrates the accuracy verification results of the proposed method for multi-band pollutant concentration detection. The horizontal axis represents the actual concentration, in milligrams per liter (mg / L), ranging from 0 to 60 mg / L, covering the entire concentration gradient from clean to heavily polluted water. The vertical axis represents the measured concentration, also in mg / L, representing the pollutant concentration values ​​measured using the proposed method. The dashed line in the figure represents the ideal fitted line—the ideal state where the measured concentration and the actual concentration are exactly equal, with a slope of 1 and an intercept of 0. The dashed-dotted line represents the raw concentration estimate, obtained by weighted summation using laboratory-calibrated band pollutant absorption sensitivity coefficients. The curve is relatively close to the ideal line in the low-concentration range (0-20 mg / L), but deviates significantly in the mid-to-high-concentration range (20-60 mg / L). In particular, above 40 mg / L, the raw estimates are generally lower than the actual concentrations, indicating systematic errors in the uncorrected detection method. The solid line represents the final pollutant concentration after correction for loss of consistency, which significantly improves detection accuracy. By applying pixel-level correction based on the wavelength of the pollutant's maximum absorption peak, the final concentration curve maintains good consistency with the ideal fitting line across the entire concentration range. The corrected detection results not only eliminate systematic bias in the original estimate but also improve the accuracy of high-concentration pollutant detection, validating the effectiveness of the proposed correction mechanism for combating consistency loss. The data distribution in the figure demonstrates that the corrected method provides stable and reliable detection results at different concentration levels.

[0066] Figure 4The adversarial consistency loss calculation process of the present invention is detailed. The nine-square grid in the upper left corner represents the spatial configuration of the target pixel and its eight-neighborhood neighborhood. The central gray square represents the target pixel to be processed, and the eight surrounding white squares represent its spatial neighboring pixels. This 3×3 neighborhood configuration is the basic unit for spatial consistency analysis, ensuring the spatial continuity and physical rationality of pixel-level correction. Starting from the target pixel and its eight-neighborhood neighborhood, the processing flow is divided into two parallel branches: the generative branch input and the ground truth branch input. The generative branch input is derived from the high-frequency residual signal output by the generator of the super-resolution hybrid generative adversarial network, which contains the spectral spatial information inferred by the network. The ground truth branch input is extracted from the corresponding region of the high-resolution remote sensing reflectance image and serves as a reference standard for the ground truth spectral spatial block. The input data of both branches is then fed into the convolutional attention network of the discriminator for processing. This network uses a deep learning algorithm to evaluate the difference between the generated content and the ground truth content, outputting a two-dimensional probability map. A global average pooling layer is used to extract the ground truth and generated probabilities for the center pixel. Simultaneously, the system calculates the spectral gradient residual and Laplacian contrast residual in parallel. The spectral gradient residual calculation module analyzes first-order differential spectral gradients across bands, extracts the difference in gradient modulus length in the central pixel, and combines this with gradient direction cosine information. The Laplace contrast residual calculation module processes the second-order Laplace spatial gradient and evaluates spatial coherence by comparing it with the neighborhood.

[0067] The following is an implementation example to illustrate the calculation process and results of the present invention at the single pixel scale. Assume that the remote sensing data source is a multispectral image acquired under clear sky conditions. The spatial resolution is increased to 5 meters per pixel after the super-resolution reconstruction in step 1. The spectral channels are selected with wavelength centers of 550 nm, 650 nm, and 860 nm, respectively. First, consider the target pixel with row index 100 and column index 200. The band reflectances in the low-resolution remote sensing reflectance image are ; After the generator inference of the super-resolution hybrid generative adversarial network, the high-frequency reflectivity residual output is ; The true value probability of water body given by the discriminator for the same pixel is .

[0068] The half-height width of the point spread function of the remote sensing imaging device in the 550-nanometer band is recorded as 3.0 meters and used as the reference maximum resolution; the half-height width of the point spread function corresponding to the three bands is 3.0 meters, 3.5 meters and 4.0 meters respectively, thus obtaining .

[0069] Reconstructing the relationship using the pixel reflectivity of the present invention , we can calculate: .

[0070] When entering step 2, laboratory water sample calibration is required to obtain the absorption sensitivity coefficients of each pollutant in the above three bands. The experimental fitting results of chemical oxygen demand and total phosphorus are set as: ; By multiplying and accumulating the high-resolution reflectance image band by band, the original estimated concentrations of chemical oxygen demand and total phosphorus can be obtained: ; ; In step 3, the adversarial consistency correction is introduced. It is assumed that the adversarial consistency loss output by the discriminator at this pixel is 0.15. The maximum absorption peak wavelength of chemical oxygen demand is set to 620 nanometers, the maximum absorption peak wavelength of total phosphorus is set to 710 nanometers, and the fixed reference wavelength is 550 nanometers. According to the correction formula of the present invention, , we can get: .

[0071] Entering step 4, you need to refer to the water quality standard limit and ecotoxicity data. Assume that the national surface water Class III limit of chemical oxygen demand is 20 mg / L, and the national surface water Class III limit of total phosphorus is 0.2 mg / L; the 96-hour median lethal concentration data in the ecotoxicity database shows that the chemical oxygen demand is 80 mg / L and the total phosphorus is 20 mg / L. . Use the comprehensive water pollution index to calculate the relationship: .

[0072] because It is obviously greater than one, which indicates that the water quality in this pixel has seriously exceeded the standard. The main source of risk is that the total phosphorus concentration far exceeds the standard limit, while the contribution of chemical oxygen demand is relatively small.

[0073] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A remote sensing water pollution detection method based on super-resolution hybrid GAN is characterized by: The method comprises: Step 1: Obtain a remote sensing reflectance image of the target area through a remote sensing imaging device; input the remote sensing reflectance image into a super-resolution hybrid generative adversarial network, and obtain a high-resolution remote sensing reflectance image through collaborative reconstruction by the generator and discriminator; Step 2: Using the experimentally calibrated band pollutant absorption sensitivity coefficient, perform weighted summation of the high-resolution remote sensing reflectance image by band to obtain the original estimated concentration value of each pollutant; Step 3: Based on the adversarial consistency loss of the super-resolution hybrid generative adversarial network and the maximum absorption peak wavelength of the pollutant, pixel-level correction is performed on the original estimated concentration of each pollutant to obtain the final concentration of each pollutant; Step 4: According to the ecotoxicity weight and water quality standard limit, the final concentration of each pollutant is weighted and normalized to generate a comprehensive water quality pollution index to represent the overall water quality pollution level.

2. The remote sensing water pollution detection method based on super-resolution hybrid GAN according to claim 1 is characterized in that: Step 1 specifically includes: obtaining a low-resolution remote sensing reflectance image, and inputting the low-resolution remote sensing reflectance image into the generator of the trained super-resolution hybrid generative adversarial network; the generator outputs a high-frequency residual signal for each pixel and each band; using the discriminator to calculate the true value probability of the water body for the same pixel, and performing confidence modulation on the high-frequency residual signal; then, according to the square of the ratio of the half-width of the point spread function of each band to the half-width of the highest resolution reference point spread function, the corresponding point spread function scale coefficient is obtained, and the modulated high-frequency residual signal is amplitude corrected using the point spread function scale coefficient; the amplitude-corrected high-frequency residual signal is summed band by band with the original low-resolution remote sensing reflectance image to obtain a super-resolution remote sensing reflectance image.

3. The remote sensing water pollution detection method based on super-resolution hybrid GAN according to claim 2 is characterized in that: In the super-resolution remote sensing reflectance image of step 1, the super-resolution reflectance corresponding to each pixel is: in, Indicates that in high-resolution remote sensing reflectance images, the row index is , the column index is The pixel has a central wavelength of Super-resolution reflectivity at ; In the remote sensing reflectance image, the row index is , the column index is The pixel has a central wavelength of Reflectivity at ; represents the high-frequency reflectivity residual predicted by the generator; The true value probability of water body calculated by the discriminator; represents the PSF scaling factor of the remote sensing imaging device; The center wavelength of the band is The point spread function of the band when ; For reference, the highest resolution.

4. The remote sensing water pollution detection method based on super-resolution hybrid GAN according to claim 3 is characterized in that: Step 2 specifically includes: for each pollutant, using laboratory standard water sample spectrum calibration to obtain the corresponding multi-band reflectance absorption sensitivity coefficient, multiplying the multi-band reflectance absorption sensitivity coefficient with the super-resolution remote sensing reflectance image item by item in the same pixel and the same band, and summing them band by band in the entire band range to obtain the original concentration estimate of each pollutant.

5. The remote sensing water pollution detection method based on super-resolution hybrid GAN according to claim 4 is characterized in that: Step 3 specifically includes: calculating the adversarial consistency loss of the super-resolution hybrid generative adversarial network within the same pixel scale, multiplying the adversarial consistency loss by the ratio of the maximum spectral absorption peak wavelength of the pollutant and the fixed reference wavelength to obtain the correction denominator; dividing the original concentration estimate by the correction denominator to obtain the final concentration of the pollutant after adversarial consistency correction.

6. The remote sensing water pollution detection method based on super-resolution hybrid GAN according to claim 5 is characterized in that: The method for calculating the adversarial consistency loss of the super-resolution hybrid generative adversarial network at the same pixel scale specifically includes: for the target pixel and its eight neighborhoods, respectively extracting the generated branch input output by the generator and the real branch input extracted from the high-resolution remote sensing reflectance image, and aligning the two one by one in terms of the number of bands and pixel positions; performing convolution degradation and downsampling on each band of the generated branch input according to the pre-calibrated corresponding band point spread function, and then realigning it with the real branch input through bicubic interpolation to obtain a physical consistency control block; synchronously sending the real branch input and the physical consistency control block to the convolutional attention network of the same discriminator, retaining only the two-dimensional probability map output by the global average pooling layer, and indexing the real probability of the center pixel and the generated probability of the center pixel; Calculate the cross-band first-order difference spectral gradient of the true branch input and the physical consistency control block separately, extract the difference in the gradient modulus of the central pixel and combine it with the gradient direction cosine to obtain the multi-band gradient residual scalar; calculate the second-order Laplace spatial gradient of the generated branch input and the true branch input, extract the difference in the Laplace value of the central pixel and compare it with the neighborhood to obtain the Laplace contrast residual, and normalize it according to the spectral reflectance variance; determine the band confidence weight, perform weighted averaging on the spectral gradient residual and the spatial coherence penalty, and jointly form a composite loss molecule with the true generation probability difference of the central pixel; calculate the local water texture scale factor based on the difference in the visible light near-infrared water body index of the target pixel and its neighborhood, perform linear normalization on the composite loss molecule, and output the adversarial consistency loss.

7. The remote sensing water pollution detection method based on super-resolution hybrid GAN according to claim 6 is characterized in that: From the high-resolution remote sensing reflectance image, the real spectral space block is extracted for the target pixel and its eight neighborhoods as the real branch input; at the same central pixel position, the inferred spectral space block of the corresponding size is synchronously extracted from the high-resolution residual signal output by the self-generator as the generation branch input; the center pixel is explicitly indexed on the two-dimensional probability map to obtain the true probability and the generated probability of the center pixel; the Laplace value difference of the center pixel is extracted and compared with the four-connected center of the neighborhood to obtain the Laplace contrast residual; the band confidence weight is determined based on the inverse of the radiation signal-to-noise ratio of each band.

8. The remote sensing water pollution detection method based on super-resolution hybrid GAN according to claim 7 is characterized in that: For row index , the column index is The comprehensive water pollution index of the pixel for: Among them, when If it is greater than 1, it means that the pollution exceeds the standard at that pixel; The most toxic of all pollutants , used to normalize the toxicity of each pollutant so that the weight ratios are all within within the scope; Indicates the The concentration of a pollutant that causes 50% mortality of a certain indicator aquatic organism within 96 hours under experimental conditions reflects the strength of its ecotoxicity. The smaller the value, the stronger the toxicity. The row index is , the column index is Adversarial consistency loss of pixels; Indicates the The upper limit of the permissible concentration of the pollutant.

Images