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

Abstract

The present application relates to the technical field of water quality detection, and more particularly to a remote sensing water quality pollution detection method based on a super-resolution hybrid GAN, which comprises the following steps: step 1: inputting a remote sensing reflectivity image into a super-resolution hybrid generative adversarial network, and obtaining a high-resolution remote sensing reflectivity image through the cooperation of a generator and a discriminator; step 2: using an experimentally calibrated band pollutant absorption sensitivity coefficient to perform weighted summation on the high-resolution remote sensing reflectivity image according to bands, and obtaining original estimated values of the concentrations of various pollutants; step 3: obtaining the final concentrations of various pollutants based on the adversarial consistency loss of the super-resolution hybrid generative adversarial network and in combination with the maximum absorption peak wavelength of the pollutants; and step 4: generating a comprehensive water quality pollution index. The present application effectively realizes the high-precision, high-credibility and high-adaptability target of pollution detection driven by remote sensing images.

Description

Technical Field

[0001] This invention belongs to the field of water quality detection technology, specifically relating to a remote sensing water pollution detection method based on super-resolution hybrid GAN. Background Technology

[0002] With the increasing prominence of water environment problems, 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 acquired through remote sensing platforms enable non-contact observation of large-scale water bodies, offering unparalleled advantages in temporal and spatial coverage compared to traditional point sampling. Utilizing the correlation between reflectance and pollutant concentration in different bands of remote sensing images, pollutant type identification, concentration estimation, and spatiotemporal variation analysis can be achieved, and this has been widely validated in various types of water bodies such as lakes, rivers, and reservoirs. However, current mainstream remote sensing water quality detection methods still have significant shortcomings in terms of accuracy, resolution, and model robustness, particularly in the collaborative identification of multiple pollutants and high-resolution spatial detail restoration, where breakthroughs are still needed.

[0003] Currently, remote sensing water quality monitoring mainly relies on a combination of multispectral imagery and empirical or semi-empirical inversion models. These methods typically use laboratory or field sampling data to perform regression modeling between the reflectance of various bands in remote sensing images and specific water quality parameters (such as chemical oxygen demand, total phosphorus, total nitrogen, chlorophyll, etc.). The most common modeling methods include multiple 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 changes in water body type, differences in lighting conditions, and significant sediment disturbance, the model's generalization ability is poor, and it is prone to overestimation or underestimation of concentrations.

[0004] Furthermore, remote sensing images are limited by the physical constraints of imaging equipment, especially when the spatial resolution of optical sensors is insufficient, making it difficult to accurately identify the distribution of fine pollutants in water bodies. Most satellite remote sensing platforms provide water quality remote sensing data with low resolution; for example, the pixel side length of medium-resolution images is typically 10 to 30 meters, making it difficult to achieve effective observation coverage in complex water systems, narrow channels, or areas with localized sewage outlets. This often results in information loss or over-smoothing in traditional models when interpolating concentration space, making it difficult to accurately characterize the boundaries of pollutant distribution. Summary of the Invention

[0005] The main objective of this invention is to provide a remote sensing water pollution detection method based on super-resolution hybrid GANs. This method reconstructs high-resolution reflectance information from remote sensing images through collaborative reconstruction of the image by a generator and discriminator, estimates concentration by combining pollutant absorption sensitivity coefficients, 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 accuracy, high reliability, and high adaptability in pollution detection driven by remote sensing images.

[0006] To solve the above problems, the technical solution of the present invention is implemented as follows:

[0007] A remote sensing method for water pollution detection based on super-resolution hybrid GANs, the method comprising:

[0008] Step 1: Acquire remote sensing reflectance images of the target area using a remote sensing imaging device; input the remote sensing reflectance images into a super-resolution hybrid generative adversarial network, and obtain high-resolution remote sensing reflectance images through collaborative reconstruction by the generator and discriminator;

[0009] Step 2: Using the experimentally calibrated band pollutant absorption sensitivity coefficients, the high-resolution remote sensing reflectance image is weighted and summed by band to obtain the original estimated concentration of each pollutant.

[0010] Step 3: Based on the adversarial consistency loss of the super-resolution hybrid generative adversarial network and combined with the wavelength of the maximum absorption peak of the pollutants, perform pixel-level correction on the original concentration estimates of each pollutant to obtain the final concentration of each pollutant;

[0011] Step 4: According to the ecotoxicity weight and water quality standard limit, the final concentrations of each pollutant are weighted and normalized to generate a comprehensive water pollution index to characterize the overall water pollution level.

[0012] Further, step 1 specifically includes: acquiring low-resolution remote sensing reflectance images and inputting them into the generator of a trained super-resolution hybrid generative adversarial network; the generator outputs high-frequency residual signals for each pixel and each band; using a discriminator to calculate the true probability of the water body based on the discrimination confidence of the same pixel, and modulating the high-frequency residual signals with confidence; then obtaining the corresponding point spread function scaling factor based on the square of the ratio of the half-width at half-maximum (WHM) of the point spread function of each band to the WHM of the highest resolution reference point spread function, and using the point spread function scaling factor to perform amplitude correction on the modulated high-frequency residual signals; and summing the amplitude-corrected high-frequency residual signals with the original low-resolution remote sensing reflectance images band by band to obtain the super-resolution remote sensing reflectance image.

[0013] Furthermore, in the super-resolution remote sensing reflectance image from step 1, the super-resolution reflectance corresponding to each pixel is:

[0014]

[0015] in, In high-resolution remote sensing reflectance imagery, the row index is... The column index is The pixel, at the center wavelength of the band Super-resolution reflectivity at that time; In remotely sensed reflectance images, the row index is... The column index is The pixel, at the center wavelength of the band Reflectance at that time; This represents the high-frequency reflectivity residual predicted by the generator; The true probability of the water body calculated for the discriminator; This indicates the PSF scaling factor of the remote sensing imaging device; The center wavelength of the band is The band point spread function at that time; For reference, the highest resolution.

[0016] Furthermore, step 2 specifically includes: for each pollutant, obtaining the corresponding multi-band reflectance absorption sensitivity coefficient using laboratory standard water sample spectral calibration, multiplying the multi-band reflectance absorption sensitivity coefficient with the super-resolution remote sensing reflectance image for each pixel and band, and summing them up band by band across the entire range to obtain the original concentration estimate of each pollutant.

[0017] 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 wavelength of the pollutant's largest spectral absorption peak to the fixed reference wavelength to obtain the correction denominator; and dividing the original concentration estimate by the correction denominator to obtain the final concentration of the pollutant after adversarial consistency correction.

[0018] Furthermore, the method for calculating the adversarial consistency loss of a super-resolution hybrid generative adversarial network within the same pixel scale specifically includes: for the target pixel and its eight neighbors, extracting the generated branch input output by the generator and the real branch input extracted from the high-resolution remote sensing reflectance image, aligning them one by one in terms of the number of bands and pixel positions; performing convolutional degradation and downsampling on each band of the generated branch input according to the pre-calibrated corresponding band point spread function, and then re-registering it with the real branch input through bicubic interpolation to obtain a physical consistency control block; simultaneously feeding 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. The algorithm calculates the cross-band first-order difference spectral gradients of the real branch input and the physical consistency control block, extracts the gradient magnitude difference of the center pixel and combines it with the gradient direction cosine to obtain the multi-band gradient residual scalar; it calculates the second-order Laplacian spatial gradients of the generated branch input and the real branch input, extracts the difference of the Laplacian value of the center pixel and compares it with the neighborhood to obtain the Laplacian contrast residual, and normalizes it according to the variance of spectral reflectance; it determines the band confidence weights, performs a weighted average of the spectral gradient residuals and spatial coherence penalty, and together with the real generation probability difference of the center pixel, constitutes the composite loss molecule; it calculates the local water texture scale factor based on the visible and near-infrared water index differences of the target pixel and its neighborhood, performs linear normalization on the composite loss molecule, and outputs the adversarial consistency loss.

[0019] Furthermore, from the high-resolution remote sensing reflectance image, a true spectral spatial block is extracted for the target pixel and its eight neighbors as the input for the true branch; at the same center pixel location, an inferred spectral spatial block of the corresponding size is simultaneously extracted from the high-resolution residual signal output by the generator as the input for the generated branch; explicit indexing of the center pixel is performed on the two-dimensional probability map to obtain the true probability and the generated probability of the center pixel; the difference in the Laplacian value of the center pixel is extracted and compared with the four-connected center of the neighborhood to obtain the Laplacian contrast residual; the confidence weight of the band is determined based on the reciprocal of the radiation signal-to-noise ratio of each band.

[0020] Furthermore, for row indexes of The column index is The comprehensive water pollution index of the pixels is:

[0021]

[0022] Among them, when A value greater than 1 indicates that the pollution level at that pixel exceeds the standard. The most toxic of all pollutants This is used to normalize the toxicity of each pollutant, ensuring that the weight ratios are all within a certain range. Within the range; Indicates the first The concentration of a pollutant that causes a 50% mortality rate in a certain indicator aquatic organism within 96 hours under experimental conditions reflects its ecotoxicity; the smaller the value, the stronger the toxicity. For row index The column index is The adversarial consistency loss of the pixels; Indicates the first The upper limit of the permissible concentration of a certain pollutant.

[0023] The remote sensing water pollution detection method based on super-resolution hybrid GAN of this invention has the following beneficial effects: This scheme can significantly improve the spatial resolution, pollutant concentration inversion accuracy, and pollution risk expression ability of remote sensing images in water quality monitoring. Compared with the traditional technical path of direct inversion based on low-resolution images, this invention effectively enhances image details and suppresses noise interference by introducing a super-resolution reconstruction mechanism that optimizes the generator and discriminator collaboratively, enabling regions such as water body boundaries, local pollution zones, and small-scale discharge outlets to be characterized with higher spatial accuracy. Furthermore, this invention uses experimentally calibrated band pollutant absorption sensitivity coefficients to achieve physical modeling of pollutant concentration, improving the interpretability and cross-regional adaptability of concentration estimation. The introduction of adversarial consistency loss in the discriminator output during concentration correction ensures that the inversion results not only conform to the pollutant reflectance characteristics in the spectral dimension but also closely approximate the real water body morphology in spatial structure, significantly reducing inversion bias caused by image reconstruction errors. In addition, this invention fully considers the differences in ecotoxicity of pollutants in the construction of the comprehensive water pollution index, effectively reflecting the actual degree of harm of pollutants through a normalized weighting mechanism, and achieving quantitative expression of pollution risk. Attached Figure Description

[0024] Figure 1 A schematic diagram of the method flow for remote sensing water pollution detection based on super-resolution hybrid GAN provided in an embodiment of the present invention;

[0025] Figure 2 This is a comparative experimental curve of super-resolution reconstruction effect provided in the embodiments of the present invention;

[0026] Figure 3 This is an experimental curve showing the multi-band pollutant concentration detection accuracy provided in an embodiment of the present invention.

[0027] Figure 4 A schematic flowchart illustrating the calculation of anti-consistency loss provided in an embodiment of the present invention. Detailed Implementation

[0028] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0029] refer to Figure 1 A remote sensing method for water pollution detection based on super-resolution hybrid GANs, the method comprising:

[0030] Step 1: Acquire remote sensing reflectance images of the target area using a remote sensing imaging device; input the remote sensing reflectance images into a super-resolution hybrid generative adversarial network, and obtain high-resolution remote sensing reflectance images through collaborative reconstruction by the generator and discriminator;

[0031] Step 2: Using the experimentally calibrated band pollutant absorption sensitivity coefficients, the high-resolution remote sensing reflectance image is weighted and summed by band to obtain the original estimated concentration of each pollutant.

[0032] Step 3: Based on the adversarial consistency loss of the super-resolution hybrid generative adversarial network and combined with the wavelength of the maximum absorption peak of the pollutants, perform pixel-level correction on the original concentration estimates of each pollutant to obtain the final concentration of each pollutant;

[0033] Step 4: According to the ecotoxicity weight and water quality standard limit, the final concentrations of each pollutant are weighted and normalized to generate a comprehensive water pollution index to characterize the overall water pollution level.

[0034] This invention discloses a remote sensing water pollution detection method based on super-resolution hybrid GAN, targeting scenarios involving multi-pollutant concentration inversion and comprehensive risk assessment. It fully integrates remote sensing image reconstruction technology with an adversarial learning framework, overcoming the limitations of traditional spectral-concentration linear mapping, which relies on single-resolution images and cannot suppress multi-scale artifacts and sensor noise. In the overall implementation process, a remote sensing reflectance image of the target area is first acquired using a remote sensing imaging device. This image undergoes radiometric calibration, geometric correction, and atmospheric correction to maintain accurate surface reflectance information. Subsequently, this 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 within its receptive domain to capture fine-grained water texture. The discriminator dynamically measures the alignment error between the generated and real objects through multi-scale perception branches. The iterative game between the two effectively recovers high-frequency details degraded by diffraction blurring, hybrid 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 point spread functions for 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.

[0035] After obtaining high-resolution remote sensing reflectance images, this invention establishes a correspondence between spectral response and water quality indicators using experimentally calibrated band-specific pollutant absorption sensitivity coefficients. The absorption sensitivity coefficients are derived from comparing the measured reflectance differences of water samples with known concentrations under laboratory conditions. Spectral fitting techniques are used to regress the characteristic absorption bands of different pollutants, simultaneously preserving wavelength dependence and spectral shape dependence. By weighted summing the absorption sensitivity coefficients and the high-resolution remote sensing reflectance images pixel-by-pixel, the original concentration estimates of each pollutant can be directly obtained. To reduce uncertainties caused by cross-sensor differences and atmospheric residuals, this invention introduces adaptive weight normalization based on band signal-to-noise ratio in the summation process and uses a water-non-water mask to filter out high-reflectance interference pixels from the land surface.

[0036] The design of the adversarial consistency loss is the key difference between this invention and existing technologies. This loss term jointly measures the difference between real and generated image blocks at the pixel level. First, it calculates the pixel authenticity deviation using the true probability and generated probability output by the discriminator. Simultaneously, it constructs an energy consistency index based on band correlation in the frequency domain to measure the fidelity of the generated result in spectral shape. Then, it comprehensively considers the ratio of the maximum absorption peak wavelength of pollutants to the reference wavelength, mapping the above two types of errors to penalty coefficients on the original concentration estimate, thus achieving pixel-level correction of the concentration of each pollutant. Since the adversarial consistency loss shares gradients with the network training process, this correction can be completed without additional computational overhead during the inference phase, enabling the concentration estimation to significantly suppress artifact diffusion while ensuring physical consistency.

[0037] In the implementation of this method, remote sensing reflectance images of the target area are first acquired using a remote sensing imaging device. These images are typically multispectral data with low spatial resolution. Due to physical constraints such as optical system resolution, sensor size, and orbital altitude, the low-resolution images cannot meet the spatial detail representation requirements for high-precision water pollution inversion. Therefore, the remote sensing reflectance images are input into a super-resolution hybrid generative adversarial network (GAN). The generator and discriminator within this network work collaboratively to enhance the details and restore the structure of the original image. Specifically, the generator predicts high-frequency compensation information from the low-resolution reflectance image, while the discriminator guides the generator to optimize the reconstruction results by comparing the pixel-level distribution and spatial consistency between the real and generated images, thereby obtaining a remote sensing reflectance image that more closely approximates the real high-resolution image.

[0038] After generating high-resolution remote sensing reflectance images, the next step is to perform preliminary estimation of pollutant concentrations. For each pollutant, its reflectance response at different spectral bands is measured using laboratory standard water samples to obtain the absorption sensitivity coefficients for each pollutant at different spectral bands. These absorption sensitivity coefficients are then weighted and summed band-by-band with the pixel reflectance values ​​of the corresponding spectral bands in the high-resolution remote sensing reflectance image to obtain the original estimated concentration of each pollutant at each pixel. This process establishes a quantitative mapping relationship from high-resolution remote sensing imagery to pollutant concentration estimation, reflecting the physical connection between remote sensing spectral characteristics and pollutant distribution.

[0039] To further improve the accuracy of pollutant concentration estimation, this method introduces adversarial consistency loss based on a super-resolution hybrid generative adversarial network as a correction factor, and combines it with the maximum absorption peak wavelength of each pollutant to perform pixel-level concentration correction. The adversarial consistency loss measures the spatial and spectral deviation between the generator output and the real high-resolution image. By comparing the pixel-level true probability with the generated probability, it reflects the degree of deviation of the current pixel in terms of realism and discriminative consistency. By combining this adversarial consistency loss with the spectral response characteristics of pollutants, a spatially sensitive adjustment mechanism for the original concentration estimate is constructed. This process emphasizes the comprehensive balance between image structure preservation and spectral response accuracy, effectively suppressing local concentration anomalies caused by remote sensing data noise, artifacts, or high-reflectivity backgrounds.

[0040] After obtaining the final concentration values ​​of each pollutant, a comprehensive water pollution index is calculated. Considering the varying toxic effects of different pollutants on aquatic organisms within the ecosystem, an ecotoxicity weighting mechanism is introduced to ensure the scientific rigor and risk orientation of the pollution assessment. Specifically, an ecotoxicity normalization factor for each pollutant is constructed based on the concentration (i.e., the median lethal concentration) that causes a 50% mortality rate in typical aquatic organisms under experimental conditions. Subsequently, the final concentration value of each pollutant is compared with its corresponding national or regional water quality standard limit to obtain the degree of exceedance. This value is then weighted and aggregated using the ecotoxicity weight to finally obtain the comprehensive water pollution index for each remote sensing pixel. This pollution index is a dimensionless parameter, facilitating regional comparison, classification, and visualization.

[0041] It is important to note that during the implementation of the method of this 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 judgment credibility must be considered. Furthermore, the absorption sensitivity coefficients of each pollutant should be established based on measured water samples, and ensure precise matching with the band response function of the remote sensing image to guarantee the physical validity of the concentration estimation process. The matching ratio between the wavelength of the pollutant's maximum absorption peak and the reference wavelength also needs to be adjusted in conjunction with the spectral configuration of the remote sensing system. To ensure high-quality output of the generated image's spectral fidelity and spatial detail, the super-resolution hybrid generative adversarial network should incorporate multiple loss functions during training, including adversarial loss, reconstruction loss, and spectral consistency loss, thereby improving the model's overall generalization ability and application stability.

[0042] Further, step 1 specifically includes: acquiring low-resolution remote sensing reflectance images and inputting them into the generator of a trained super-resolution hybrid generative adversarial network; the generator outputs high-frequency residual signals for each pixel and each band; using a discriminator to calculate the true probability of the water body based on the discrimination confidence of the same pixel, and modulating the high-frequency residual signals with confidence; then obtaining the corresponding point spread function scaling factor based on the square of the ratio of the half-width at half-maximum (WHM) of the point spread function of each band to the WHM of the highest resolution reference point spread function, and using the point spread function scaling factor to perform amplitude correction on the modulated high-frequency residual signals; and summing the amplitude-corrected high-frequency residual signals with the original low-resolution remote sensing reflectance images band by band to obtain the super-resolution remote sensing reflectance image.

[0043] First, a low-resolution remote sensing reflectance image covering the target area is acquired from a remote sensing imaging device. This image is a collection of multi-band two-dimensional images, typically radiometrically calibrated in reflectance units. Due to the limited spatial resolution of this image, it is significantly insufficient in representing the spatial distribution of water pollution. Therefore, this low-resolution remote sensing reflectance image needs to be input into a pre-trained super-resolution hybrid generative adversarial network generator. The generator employs a deep convolutional network structure, capable of learning the spatial distribution between pixels and the spectral correlation between bands from the original image, and outputting high-frequency residual signals for each pixel and its corresponding band. This high-frequency residual signal represents the generator's reconstruction of the details lost in the original image due to point spread function blurring, including important features such as texture edges, local contrast, and inter-band reflectance jumps.

[0044] Subsequently, the high-frequency residual signal output by the generator is fused with the discriminator's calculation results. The discriminator is a discriminative subnetwork in a super-resolution hybrid generative adversarial network, capable of identifying local differences between real remote sensing images and generated images. For each pixel, a recognition model based on spatial neighborhood and spectral features is established within the discriminator, outputting the true confidence score of the water pixel, i.e., the ground truth probability of the water body. This ground truth probability of the water body is used to modulate the confidence score of the high-frequency residual signal output by the generator. Specifically, this involves multiplying the original residual signal by the probability result output by the discriminator, thereby enhancing features in reliable regions and suppressing features in unreliable regions, thus improving the physical consistency and content stability of the reconstructed image.

[0045] In the actual amplitude recovery process, the modulated high-frequency residual signal needs to be scaled and normalized using the point spread function (FWHM) of the remote sensing imaging system. Because remote sensing imaging systems have different optical response characteristics in different bands, and the spatial resolution of each band varies due to factors such as lens diffraction, filter thickness, and sensor structure, normalization is required based on the FWHM of the point spread function for each band and the FWHM of the system's highest-resolution reference point spread function. The square of the ratio of these two values ​​is used as a scale conversion factor to construct the point spread function scaling coefficient. This coefficient is used to correct the amplitude of the high-frequency residual signal, ensuring that the final reconstructed high-resolution information is consistent with the actual observation conditions of each band in terms of geometric scale and radiation amplitude, preventing inconsistencies in reconstructed information or spatial resolution mismatches between different bands.

[0046] After amplitude correction, 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 details in the original image, ultimately generating a set of super-resolution remote sensing reflectance images with higher spatial resolution, richer texture structure, and stronger inter-band consistency. This provides a more reliable data foundation for subsequent fine identification and concentration inversion of water pollutants. In the above implementation process, special attention must be paid to ensuring the stability of the parameter update mechanism between the generator and the discriminator to avoid mode collapse or gradient oscillation during training. Simultaneously, the true probability of the water body output by the discriminator needs to be regularized to ensure its range is limited to a reasonable interval. The measurement of the half-width at half-maximum (WHM) of the point spread function should be combined with the specific model of the remote sensing imaging device and experimental calibration results to ensure that the point spread function scaling factor is physically usable.

[0047] Furthermore, in the super-resolution remote sensing reflectance image from step 1, the super-resolution reflectance corresponding to each pixel is:

[0048]

[0049] in, In high-resolution remote sensing reflectance imagery, the row index is... The column index is The pixel, at the center wavelength of the band Super-resolution reflectivity at that time; In remotely sensed reflectance images, the row index is... The column index is The pixel, at the center wavelength of the band Reflectance at that time; This represents the high-frequency reflectivity residual predicted by the generator; The true probability of the water body calculated for the discriminator; This indicates the PSF scaling factor of the remote sensing imaging device; The center wavelength of the band is The band point spread function at that time; For reference, the highest resolution.

[0050] This relationship explicitly includes the original low-resolution remote sensing reflectance image. High-frequency reflectivity residuals from generator output The values ​​are superimposed, and the true probability of the water body from the discriminator is introduced. As a confidence modulation factor, and through This diffusion function scaling factor achieves amplitude correction for band resolution differences. Its working mechanism lies in the generator first predicting based on the local spatial context of low-resolution pixels and cross-band spectral modes. The residual numerically represents the magnitude of detail lost due to optical diffraction blurring; subsequently, the discriminator assigns a value to the same pixel. This value, after being mapped using the Sigmoid form, falls between zero and one, and can be considered as the confidence weight of a pixel in the true remote sensing statistical distribution. Therefore, the denominator in the formula... This constitutes an adaptive suppression mechanism for high-frequency residual signals, enabling the generator to fully preserve details in the high-confidence region and automatically converge to zero in the low-confidence region, thus avoiding artifact spread.

[0051] On the other hand, differences in lens diffraction limit, filter refractive index, and detector spacing lead to variations in different wavelength bands. Since they are different from each other, to ensure that the recovered detail amplitudes of each band have a uniform physical scale, this invention measures the half-width at half-maximum (FWHM) of the point spread function of the reference highest resolution band. With the target band half width at half maximum (FWHM) The square ratio is introduced This factor numerically acts as an amplitude scaling factor for the high-frequency residual signal, compensating for detail loss caused by diffraction in the long band and suppressing potential over-sharpening in the short band, thereby ensuring... Consistency of spatial resolution across dimensions. In actual deployment, the generator and discriminator need to be updated alternately on the same training image set. The generator's loss function, in addition to adversarial loss, also needs to incorporate spectral consistency constraints, while the discriminator needs to be trained using the joint distribution of real and synthetic images to ensure... It reflects the probability of pixel authenticity on a global scale.

[0052] The full width at half maximum (FWHM) parameter of the point spread function needs to be obtained through precise optical testing in a laboratory or through observations of stars in orbit, and must be tabulated and stored during image preprocessing to ensure accuracy. The calculations are accurate and reliable. The entire super-resolution reconstruction process uses pixels as the processing unit, maintaining spatial topology invariance and avoiding errors introduced by coordinate interpolation; simultaneously, each band is applied independently during the generation process. The scaling factor ensures that the radiometric calibration relationship between spectral channels is not disrupted. It is important to note that in highly reflective bright spots or shadowed areas, the true probability of the water body output by the discriminator may show oversaturated or undersaturated values. This invention adjusts the slope of the sigmoid curve by incorporating a temperature coefficient during the training phase to achieve this. The distribution is smoother to prevent convergence instability caused by local gradient explosion. Furthermore, to avoid spectral jumps caused by excessive differences in point spread functions across different bands, this invention adds a bandpass filter kernel at the high-frequency residual signal output to limit the highest usable frequency of the predicted content, ensuring it does not exceed... The maximum supported resolution. Through the methods described above, the final output... In terms of spatial detail, spectral morphology, and radiation amplitude, it is consistent with the actual capabilities of remote sensing imaging hardware, providing a high-quality data foundation for subsequent pollutant concentration estimation, and significantly improving the stability, accuracy, and physical reliability of the entire remote sensing water pollution detection process.

[0053] Furthermore, step 2 specifically includes: for each pollutant, obtaining the corresponding multi-band reflectance absorption sensitivity coefficient using laboratory standard water sample spectral calibration, multiplying the multi-band reflectance absorption sensitivity coefficient with the super-resolution remote sensing reflectance image for each pixel and band, and summing them up band by band across the entire range to obtain the original concentration estimate of each pollutant.

[0054] First, spectral measurements of laboratory standard water samples are required for each pollutant. Specifically, standard water samples with a certain concentration gradient are collected, and the reflectance variation trend at the center wavelength of multiple remote sensing bands is recorded using a spectral measuring instrument. Combined with the known water sample concentration, methods such as linear fitting, principal component regression, or partial least squares are employed to determine the absorption sensitivity coefficient between pollutant concentration and reflectance at each band. This absorption sensitivity coefficient reflects the proportional relationship between the spectral absorption intensity of a specific pollutant at each band and its concentration change, and has a clearly defined mass concentration unit, typically milligrams per liter. The absorption sensitivity coefficient is an empirical modeling result of pollutants under specific spectral response conditions, exhibiting band dependence and pollutant type dependence. Different pollutants show significant differences in absorption characteristics at different bands; therefore, experimental calibration is required for each pollutant to establish a set of band-pollutant corresponding absorption sensitivity coefficient matrices.

[0055] The absorption sensitivity coefficient matrix described above is fused with the super-resolution remote sensing reflectance image generated in step 1, ensuring a one-to-one correspondence between the data's spatial location and spectral location. Specifically, for each pixel location in the image and its corresponding spectral reflectance values, the absorption sensitivity coefficient of the pollutant is extracted item by item, and then multiplied item by item with the reflectance value of that spectral band. This process is a pixel-by-pixel, band-by-band calculation, and the intermediate results obtained represent the quantified spectral response of the pollutant to its concentration in each spectral band. Subsequently, a band-by-band summation operation is performed on the response results of all bands to accumulate and obtain the original estimated concentration of the target pollutant at that pixel.

[0056] In practice, to ensure consistency between the multi-band reflectance absorption sensitivity coefficient and the super-resolution remote sensing reflectance image, it is necessary to ensure that the center wavelength, bandwidth, and sampling resolution of the spectral bands of the remote sensing image are consistent with the parameters of the spectral device used in the experimental calibration, or consistent after interpolation and normalization. Furthermore, the super-resolution remote sensing reflectance image should have undergone atmospheric correction and radiometric calibration, and the reflectance values ​​of all bands should be the apparent reflectance or surface reflectance of the ground object, avoiding interference from external factors such as atmospheric path radiation and changes in solar altitude angle. For the extraction of the absorption sensitivity coefficient, a sample system under multiple representative water body scenarios should be established, considering the influence of variables such as water color type, water transparency, and suspended solids content on the spectral response, to avoid systematic biases in concentration estimation results under different environmental conditions due to insufficient model generalization.

[0057] 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 band selection, introducing nonlinear transformations, or utilizing auxiliary parameters. Furthermore, when some bands suffer from saturation or low signal-to-noise ratio, robustness should be improved by setting band masking or dynamically adjusting band weights. All original concentration estimates are kept in the same mass concentration units as the absorption sensitivity coefficient and are retained in subsequent steps as the basis for the final pollutant concentration estimation.

[0058] 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 wavelength of the pollutant's largest spectral absorption peak to the fixed reference wavelength to obtain the correction denominator; and dividing the original concentration estimate by the correction denominator to obtain the final concentration of the pollutant after adversarial consistency correction.

[0059] First, using pixels as processing units, the original concentration estimate of each pixel in the super-resolution remote sensing reflectance image is extracted. Then, the consistency evaluation result between the true probability of water body output by the discriminator at that pixel location and the generated image is used to construct the adversarial consistency loss for that pixel. This loss value can be regarded as an indicator of the deviation between the generated image and the real remote sensing image at that location in terms of spatial structure and spectral morphology. The larger the value, the more likely there are problems such as detail deviation, insufficient discrimination confidence, or spatial artifact superposition during the super-resolution restoration process. Therefore, this loss term is used as a suppression factor in concentration inversion to adjust the original concentration estimate in uncertain regions and prevent abnormal amplification of pollutant concentration.

[0060] After calculating the anti-consistency loss, to further enhance its physical rationality in the spectral dimension, this invention multiplies the loss term by the ratio between the maximum absorption peak wavelength of the pollutant within the remote sensing band and a fixed reference wavelength, forming a correction denominator with spectral dependence. The maximum absorption peak wavelength refers to the center wavelength where the pollutant absorbs most strongly in the actual water body spectrum, usually obtained by laboratory spectral measurements, representing the pollutant's sensitivity to spectral information in 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, serving as a band normalization reference to construct a unified scale factor. This ratio is used to modulate the anti-consistency loss through band response, allowing different pollutants to obtain different correction factors based on their spectral activity, thus enhancing the pollutant specificity and spectral dependence of the correction process.

[0061] Subsequently, the original concentration estimate is divided by the constructed correction denominator to obtain the final concentration value of the corresponding pollutant in that pixel. This concentration value is the corrected result after spatial consistency penalty and spectral sensitivity adjustment, and numerically reflects the physical laws of pollutant distribution in real water bodies. This correction mechanism effectively avoids concentration anomalies caused by high-frequency detail distortion, local structure shift, or overfitting of the inversion model, improving the spatial smoothness and physical interpretability of the concentration estimate.

[0062] During implementation, the adversarial consistency loss should be calculated by fusing differences in image spatial features and spectral structural gradients to ensure simultaneous spatial and spectral discrimination capabilities. The selection of the wavelength of the maximum absorption peak must be based on the spectral characteristics of the pollutants and the distribution range of the remote sensing bands, ensuring that this ratio falls within a reasonable dynamic range to avoid excessive amplification or compression errors in the final concentration. Furthermore, to prevent the discriminator from outputting abnormally high misclassification probabilities in edge or high-reflectivity regions, background masking and water body prior enhancement mechanisms can be introduced during the training phase to optimize the discriminator's consistency response performance within real water body areas.

[0063] By implementing this step, the present invention can achieve pixel-level pollutant concentration correction, construct a high-precision concentration estimation mechanism for super-resolution remote sensing data, significantly improve the reliability and distribution continuity of pollutant concentration inversion, and provide more reliable basic data for subsequent pollution index evaluation. The entire process, while maintaining the original physical structure of the concentration, introduces the discriminative power of a generative adversarial network, achieving a closed loop of physical consistency from image data to pollution concentration results in remote sensing inversion.

[0064] Furthermore, the method for calculating the adversarial consistency loss of a super-resolution hybrid generative adversarial network within the same pixel scale specifically includes: for the target pixel and its eight neighbors, extracting the generated branch input output by the generator and the real branch input extracted from the high-resolution remote sensing reflectance image, aligning them one by one in terms of the number of bands and pixel positions; performing convolutional degradation and downsampling on each band of the generated branch input according to the pre-calibrated corresponding band point spread function, and then re-registering it with the real branch input through bicubic interpolation to obtain a physical consistency control block; simultaneously feeding 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. The algorithm calculates the cross-band first-order difference spectral gradients of the real branch input and the physical consistency control block, extracts the gradient magnitude difference of the center pixel and combines it with the gradient direction cosine to obtain the multi-band gradient residual scalar; it calculates the second-order Laplacian spatial gradients of the generated branch input and the real branch input, extracts the difference of the Laplacian value of the center pixel and compares it with the neighborhood to obtain the Laplacian contrast residual, and normalizes it according to the variance of spectral reflectance; it determines the band confidence weights, performs a weighted average of the spectral gradient residuals and spatial coherence penalty, and together with the real generation probability difference of the center pixel, constitutes the composite loss molecule; it calculates the local water texture scale factor based on the visible and near-infrared water index differences of the target pixel and its neighborhood, performs linear normalization on the composite loss molecule, and outputs the adversarial consistency loss.

[0065] In the specific implementation process, firstly, a two-dimensional neighborhood window consisting of nine pixels in three rows and three columns is extracted from the high-resolution remote sensing reflectance image, centered on the target pixel to be analyzed. This neighborhood window is simultaneously registered in the generator output image and the reference ground truth image, serving as the input for the generation branch and the ground truth branch, respectively. To ensure the effectiveness of the comparison and the alignment of the spatial structure, the generation branch and the ground truth branch must correspond completely one-to-one in terms of the number of bands and pixel positions; that is, all band channels must be consistent, and spatial slicing must be performed using the same row and column indices. The above inputs constitute the spatial-spectral joint pixel block structure required in the generation and discrimination path, which is the input basis for the entire loss calculation.

[0066] To simulate the point spread phenomenon in different bands of actual remote sensing imaging systems and further construct a physical degradation comparison of the generated image, after obtaining the input of the generation branch, a two-dimensional convolution operation needs to be performed on each band according to a pre-calibrated point spread function. The width of the convolution kernel is based on the wavelength corresponding to the center wavelength of the band. Setup. 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 it is completely identical in size to the true branch input, the downsampled image undergoes bicubic interpolation reconstruction to restore it to the original high-resolution pixel scale, ultimately yielding a physically consistent control block. This physically consistent control block is the restored result of the generated image after being processed by the real remote sensing imaging degradation model; its structure will be used to compare various feature differences with the true branch image.

[0067] The aforementioned real branch input and the physically consistent control block are synchronously input into the discriminator convolutional attention network with shared weights. Standard convolution operations and multi-head attention aggregation are performed, and a two-dimensional probability map output from the global average pooling layer is retained at the end. This map is used to quantify the confidence probability distribution that the input region is a real image. On this probability map, the true probability and generated probability of the center pixel are extracted using the index of the center pixel, obtaining the discriminative difference of the current pixel under the discriminator's semantics. This probability difference serves as one of the loss components of the main discriminative path, representing the degree of semantic difference between the generated result and the real image.

[0068] To enhance the responsiveness to spectral structure information, the first-order difference spectral gradients of the real branch input and the physically consistent control block need to be calculated across all band dimensions, and the gradient magnitude of the central pixel needs to be calculated. Furthermore, the direction cosine between the two gradient vectors is extracted, and the weighted residual value of the spectral direction difference is calculated based on the magnitude difference, forming a multi-band gradient residual scalar. This residual reflects the degree of detail distortion in the spectral morphology of the generated image and plays a crucial role in the spectral unmixing of water pollutants.

[0069] To further consider spatial structure fidelity, a second-order Laplacian operator convolution is performed simultaneously on the real branch input and the generated branch input to obtain the spatial edge response map, and the difference in Laplacian values ​​of the central pixel is extracted. This difference, together with the Laplacian responses of surrounding neighboring pixels, forms a local contrast difference index, reflecting the ability of the generated image to preserve structural details; this is called the Laplacian contrast residual. Considering the differences in brightness levels and signal-to-noise ratios across different bands in remote sensing images, this residual needs to be normalized according to the spectral reflectance variance to ensure a uniform discrimination standard across different bands and suppress the error amplification effect of high-reflectance or high-saturation bands.

[0070] After constructing all discrimination 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 radiation signal-to-noise ratio (SNR) of each band; bands with lower SNR receive lower weights to reduce the impact of low-quality bands on the final loss value. The spectral gradient residuals and spatial coherence penalty are weighted and averaged using these confidence weights, and then combined with the difference in the true generation probability of the center pixel to form a composite loss molecule.

[0071] Finally, to accommodate the complexity of texture structures in different types of water bodies, this invention constructs a local water body texture scale factor based on the differences in visible and near-infrared water body indices of the target pixel and its eight neighboring regions. This factor reflects the texture complexity or water color variation level of the water body in that region, and accordingly performs linear normalization on the aforementioned composite loss molecule to suppress residual anomalies caused by complex water body regions. The normalized result is the final anti-consistency loss, used in the subsequent pixel-level correction process for pollutant concentration.

[0072] Furthermore, from the high-resolution remote sensing reflectance image, a true spectral spatial block is extracted for the target pixel and its eight neighbors as the input for the true branch; at the same center pixel location, an inferred spectral spatial block of the corresponding size is simultaneously extracted from the high-resolution residual signal output by the generator as the input for the generated branch; explicit indexing of the center pixel is performed on the two-dimensional probability map to obtain the true probability and the generated probability of the center pixel; the difference in the Laplacian value of the center pixel is extracted and compared with the four-connected center of the neighborhood to obtain the Laplacian contrast residual; the confidence weight of the band is determined based on the reciprocal of the radiation signal-to-noise ratio of each band.

[0073] First, from the high-resolution remote sensing reflectance image, a true spectral spatial block is extracted from the target pixel and its eight neighbors. This spectral spatial block has a spatial dimension of three rows and three columns, containing nine pixels, with the spectral dimension corresponding to the total number of bands contained in the remote sensing image. This true spectral spatial block serves as the input to the true branch, preserving the local spectral structure and spatial distribution characteristics of the target region, providing the original input for the subsequent true image discrimination path. Simultaneously, at the same central pixel location—that is, at a position strictly aligned with the spatial index of the true spectral spatial block—an inferred spectral spatial block with the same size as the true branch is extracted from the high-resolution residual signal output by the generator, serving as the input to the generated branch. This generated branch input preserves the local spatial texture and band structure information predicted by the generator during reconstruction, and together with the true branch, forms the input basis for the discriminator's dual-branch path.

[0074] The inputs from the two branches mentioned above are simultaneously fed into the convolutional attention structure of the super-resolution hybrid generative adversarial network discriminator. After multiple layers of spatial convolution, attention feature aggregation, and feature mapping, a two-dimensional probability map is finally output by a global average pooling layer. Each pixel value in this probability map represents the probability that the corresponding pixel is judged as a real remote sensing pixel under the discriminator's semantic recognition. To obtain a quantitative indicator for subsequent residual construction, this invention performs an explicit indexing operation on the center pixel of the two-dimensional probability map, directly extracting the true probability and generation probability of the center pixel under the true branch path and the generation branch path, respectively denoted as the true probability and generation probability of the center pixel. The difference between these two probability values ​​is used to characterize the degree of deviation of the statistical realism of the generated image at that pixel, and is one of the inputs to the core discriminative path in the adversarial consistency loss.

[0075] To further enhance sensitivity to changes in spatial structure, this invention introduces higher-order spatial gradient response features within the central pixel and its neighborhood. Specifically, second-order Laplacian spatial gradient maps are calculated for both the true branch input and the generated branch input. These Laplacian spatial gradient maps capture edge responses and high-frequency structural changes in the image, serving as a crucial basis for judging detail consistency. After extracting the Laplacian response value of the central pixel from this map, it is further compared and analyzed with the Laplacian values ​​of the four-connected neighboring central pixels to construct a local contrast residual. This residual, called the Laplacian contrast residual, quantifies the difference in the degree to which the generated image preserves the structural edges around the pixel. A larger residual value indicates a poorer restoration of edge details in the generated image at that pixel, requiring increased weight constraints in subsequent corrections.

[0076] To ensure a unified weighting system for the fusion of various residuals across the spectral dimension, and to suppress interference from bands with low signal-to-noise ratios (SNR) on the residual components, this invention calculates band confidence weights based on the radiation SNR of each band. Specifically, the SNR of each band across multiple training samples is calculated, and its reciprocal is used as a weight component; bands with higher SNRs are assigned larger confidence weights. These band confidence weights are used in subsequent weighted operations for the fusion of multi-band gradient residuals, spatial contrast residuals, and center pixel probability differences, ensuring that the final output adversarial consistency loss exhibits reasonable sensitivity and stability across multiple bands and structural levels.

[0077] Furthermore, for row indexes of The column index is The comprehensive water pollution index of the pixels is:

[0078]

[0079] Among them, when A value greater than 1 indicates that the pollution level at that pixel exceeds the standard. The most toxic of all pollutants This is used to normalize the toxicity of each pollutant, ensuring that the weight ratios are all within a certain range. Within the range; Indicates the first The concentration of a pollutant that causes a 50% mortality rate in a certain indicator aquatic organism within 96 hours under experimental conditions reflects its ecotoxicity; the smaller the value, the stronger the toxicity. For row index The column index is The adversarial consistency loss of the pixels; Indicates the first The upper limit of the permissible concentration of a certain pollutant.

[0080] In the implementation process, firstly, after the pollutant concentration correction step, a concentration matrix is ​​constructed according to the pollutant dictionary index order, ensuring that all concentration values ​​uniformly use milligrams per liter as the unit of mass concentration. Subsequently, an ecotoxicology database is searched to obtain a set of 96-hour median lethal concentration (LD50) data consistent with the pollutant types monitored by remote sensing; the minimum value is recorded as [value missing]. At the same time, read the corresponding To avoid computational distortion caused by differences in numerical scales, it is necessary to... Perform a missing value check. If no publicly available data is available for a pollutant, a reference value with a similar chemical structure or ecotoxicity parameter can be used, and the source should be recorded in the documentation. Then, based on applicable national or regional water environment standards, determine the upper limit of permissible concentration for each pollutant. During the data preparation phase, concentration units should be standardized to ensure consistency. and The same as milligrams per liter.

[0081] The calculation phase is performed by traversing the pixels: for each pixel position Call the saved Compared with standard limits The ratio of exceeding the standard is then divided into the toxicity weighting ratio. Multiply by each pollutant to calculate the risk score for each pollutant, and finally sum the risk scores for all pollutants to obtain the final result. It is recommended to use single-precision floating-point operations during implementation to reduce the storage pressure caused by large-scale cell loops; if deployed on a low-power platform, memory usage can be reduced through block computation and cache reuse strategies. To ensure... For interpretability, a corresponding hierarchical threshold table should be established, for example, […]. Classified as a safe level, 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.

[0082] In terms of precision control, attention should be paid to The range of ratios needs to be controlled. Excessive differences in ecotoxicity may lead to an extreme amplification of the weight of a particular pollutant, causing instability in the index value. An upper threshold can be set for the comparison values, for example, limiting them to within ten. For values ​​exceeding ten, a percentile truncation strategy can be used to maintain the robustness of the index results. Additionally, The acquisition of the index 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 specular reflection. It is recommended to perform cloud masking and water body masking on the relevant pixels before concentration correction to avoid non-target areas interfering with the accuracy of the index.

[0083] In software implementation, The computation module can be encapsulated as a matrix operation interface, processing all pixels at once through a vectorized programming environment, thus improving operational efficiency. At the hardware implementation level, it can be combined with graphics processing units or tensor processing units for parallel computation, meeting the needs of real-time pollution index generation for large-scale remote sensing imagery. The final result... The raster layer can serve as a base map for visualizing pollution distribution, or it can be input into downstream water environment models or river and lake health assessment systems to provide a basis for water quality improvement measures.

[0084] Figure 2 This paper presents a comparison of the reconstruction effects of the super-resolution hybrid generative adversarial network (GAN) on remote sensing reflectance images, as described in this invention. The horizontal axis represents wavelength in nanometers, covering the visible to near-infrared spectral range from 400 nm to 900 nm, encompassing the main spectral absorption characteristic bands of water pollutants. The vertical axis represents reflectance, ranging from 0.0 to 1.0, characterizing the spectral reflectance intensity of the water surface at different wavelengths. The figure includes two contrasting 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 curve comparison clearly shows that the low-resolution remote sensing reflectance image exhibits relatively smooth reflectance changes across all bands, with significant loss of spectral detail, especially in the 500-700 nm band, where the curve shows 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 detail variations, with more pronounced absorption and reflection peaks in the pollutant-sensitive bands. Particularly in the 400-500 nm and 600-800 nm wavelength ranges, the high-resolution curves exhibit sharper spectral response changes, and these detailed features are crucial for subsequent quantitative inversion of pollutant concentrations. Through collaborative reconstruction by the generator and discriminator, high-resolution remote sensing reflectance images effectively preserve high-frequency information of the water body spectrum, providing a reliable data foundation for accurately detecting water pollutant concentrations.

[0085] Figure 3 This paper presents the accuracy verification results of the method of this invention in 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 complete concentration gradient from clean water to heavily polluted water. The vertical axis represents the detected concentration, also in mg / L, indicating the pollutant concentration value measured by the method of this invention. The dashed line in the figure represents the ideal fitted straight line, i.e., the ideal state where the detected concentration is exactly equal to the actual concentration. This line has a slope of 1 and an intercept of 0. The dotted line represents the original concentration estimate obtained by weighted summation using the band pollutant absorption sensitivity coefficients calibrated in the laboratory. It can be observed that this curve is relatively close to the ideal straight line in the low concentration range (0-20 mg / L), but deviates significantly in the medium-to-high concentration range (20-60 mg / L), especially above 40 mg / L, where the original estimate is generally lower than the actual concentration, indicating that the uncorrected detection method has systematic errors. The solid line represents the final pollutant concentration after correction for anti-consistency loss, and this curve significantly improves the detection accuracy. By incorporating the wavelength of the pollutant's maximum absorption peak for pixel-level correction, the final concentration curve maintains good consistency with the ideal fitted line across the entire concentration range. The corrected detection results not only eliminate systematic biases in the original estimate but also improve the detection accuracy of high-concentration pollutants, validating the effectiveness of the anti-consistency loss correction mechanism in this invention. The data point distribution in the figure shows that the corrected method provides stable and reliable detection results at different concentration levels.

[0086] Figure 4The adversarial consistency loss calculation process of this invention is illustrated in detail. The nine-square grid diagram in the upper left corner represents the spatial configuration of the target pixel and its eight neighbors, where the central gray square represents the target pixel to be processed, and the surrounding eight white squares are its spatial neighbor 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 neighbors, the processing flow is divided into two parallel branches: the generation branch input and the ground truth branch input. The generation branch input comes from the high-frequency residual signal output by the super-resolution hybrid generative adversarial network generator, containing 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, serving as the reference standard for the ground truth spectral spatial block. The input data of the two branches are then fed into the convolutional attention network of the discriminator for processing. This network evaluates the difference between the generated content and the ground truth content through a deep learning algorithm, outputs a two-dimensional probability map, and extracts the ground truth probability and the generated probability of the center pixel through a global average pooling layer. Simultaneously, the system calculates the spectral gradient residual and the Laplacian contrast residual in parallel. The spectral gradient residual calculation module analyzes the first-order difference spectral gradient across bands, extracts the gradient magnitude difference of the center pixel, and combines it with the gradient direction cosine information. The Laplacian contrast residual calculation module processes the second-order Laplacian spatial gradient, obtaining spatial coherence assessment results by comparing it with the neighborhood.

[0087] The following is an implementation example illustrating the calculation process and results of this invention at the single-pixel scale. Assume the remote sensing data source is a multispectral image acquired under clear sky conditions. After super-resolution reconstruction in step 1, the spatial resolution is improved to 5 meters per pixel. The spectral channels are selected with wavelength centers of 550 nm, 650 nm, and 860 nm. 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 as follows: After generator inference via a super-resolution hybrid generative adversarial network, the high-frequency reflectivity residual output is: The discriminator gives the true probability of water bodies for the same pixel as follows: .

[0088] The half-width at half-maximum (WHM) of the point spread function (PSF) of the remote sensing imaging device in the 550 nm band is denoted as 3.0 m, and is used as the reference highest resolution; the WHMs of the PSF corresponding to the three bands are 3.0 m, 3.5 m, and 4.0 m, respectively, from which we obtain... .

[0089] Pixel reflectance reconstruction relationship using the present invention The calculation yields:

[0090] .

[0091] In step 2, laboratory water samples need to be calibrated to obtain the absorption sensitivity coefficients of each pollutant in the three wavelength bands mentioned above. The experimental fitting results for chemical oxygen demand (COD) and total phosphorus are set as follows:

[0092] ;

[0093] By multiplying and summing the high-resolution reflectance images band by band, the original estimates of chemical oxygen demand (COD) and total phosphorus concentrations can be obtained:

[0094] ;

[0095] ;

[0096] Step 3 introduces adversarial consistency correction, assuming 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 nm, the maximum absorption peak wavelength of total phosphorus is set to 710 nm, and a fixed reference wavelength is taken as 550 nm. This is based on the correction formula of the present invention. We can obtain:

[0097] .

[0098] Step 4 requires referencing water quality standard limits and ecotoxicity data. Let's assume the national limit for Chemical Oxygen Demand (COD) in Class III surface water is 20 mg / L, and the national limit for Total Phosphorus (TP) in Class III surface water is 0.2 mg / L; the 96-hour median lethal concentration (LD50) data in the ecotoxicity database shows COD at 80 mg / L and TP at 20 mg / L. Therefore... The relationship is calculated using the comprehensive water pollution index:

[0099] .

[0100] because If the value is significantly greater than 1, the water quality at that pixel is considered to be seriously exceeding the standard. The main source of risk is the total phosphorus concentration, which is far exceeding the standard limit, while the contribution of chemical oxygen demand is relatively small.

[0101] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate 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 a super-resolution hybrid GAN, characterized in that, The method comprises: Step 1: obtaining remote sensing reflectivity images of a target area by a remote sensing imaging device; inputting the remote sensing reflectivity images into a super-resolution hybrid generative adversarial network, and reconstructing a high-resolution remote sensing reflectivity image by a generator and a discriminator in cooperation; Step 2: weighting and summing the high-resolution remote sensing reflectivity images according to the band pollutant absorption sensitivity coefficients calibrated by experiments to obtain original concentration estimates of each pollutant; Step 3: performing pixel-level correction on the original concentration estimates of each pollutant based on the adversarial consistency loss of the super-resolution hybrid generative adversarial network and in combination with the maximum absorption peak wavelength of the pollutant to obtain the final concentration of each pollutant; Step 4: weighting and normalizing the final concentration of each pollutant according to the ecological toxicity weight and the water quality standard limit value, and generating a comprehensive water quality pollution index to represent the overall water pollution degree; The super-resolution reflectivity corresponding to each pixel in the super-resolution remote sensing reflectivity image of step 1 is: wherein, represents a super-resolution reflectance of a pixel with row index and column index in a high-resolution remote sensing reflectance image at a band center wavelength ; is a reflectance of a pixel with row index and column index in a remote sensing reflectance image at a band center wavelength ; represents a generator-predicted high-frequency reflectance residual; is a water-body ground-truth probability computed by a discriminator; represents a PSF scaling factor of a remote sensing imaging device; is a band point spread function at a band center wavelength ; is a reference highest resolution; Step 3 specifically comprises: 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 to the fixed reference wavelength to obtain a correction denominator, and dividing the original concentration estimate by the correction denominator to obtain the final concentration of the pollutant after adversarial consistency correction; For a pixel with row index and column index , the comprehensive water pollution index is: Among them, when A value greater than 1 indicates that the pollution level at that pixel exceeds the standard. The most toxic of all pollutants This is used to normalize the toxicity of each pollutant, ensuring that the weight ratios are all within a certain range. Within the range; Indicates the first The concentration of a pollutant that causes a 50% mortality rate in a certain indicator aquatic organism within 96 hours under experimental conditions reflects its ecotoxicity; the smaller the value, the stronger the toxicity. For row index The column index is The adversarial consistency loss of the pixels; Indicates the first The upper limit of the permissible concentration of a certain pollutant.

2. The remote sensing water quality pollution detection method based on super-resolution hybrid GAN of claim 1, wherein, Step 1 specifically comprises: obtaining a low-resolution remote sensing reflectivity image, and inputting the low-resolution remote sensing reflectivity image into the generator of the trained super-resolution hybrid generative adversarial network; the generator outputs high-frequency residual signals for each pixel and each band; the discriminator calculates the water body true value probability by the discrimination confidence of the same pixel, and modulates the high-frequency residual signals by the confidence; then, the point spread function scale coefficient is obtained according to the square of the ratio of the half width of each band point spread function to the half width of the highest resolution reference point spread function, and the amplitude of the modulated high-frequency residual signals is corrected using the point spread function scale coefficient; the amplitude-corrected high-frequency residual signals and the original low-resolution remote sensing reflectivity image are added wave by wave to obtain the super-resolution remote sensing reflectivity image. 3.The remote sensing water quality pollution detection method based on super-resolution hybrid GAN of claim 2, wherein, Step 2 specifically comprises: for each pollutant, the corresponding multi-band reflectivity absorption sensitivity coefficient is obtained by laboratory standard water sample spectral calibration, the multi-band reflectivity absorption sensitivity coefficient is multiplied with the super-resolution remote sensing reflectivity image by wave by wave within the same pixel and the same band, and the original concentration estimate of each pollutant is obtained by summing all wave bands. 4.The remote sensing water quality pollution detection method based on super-resolution hybrid GAN of claim 1, wherein, The method for calculating the adversarial consistency loss of the super-resolution hybrid generative adversarial network in the same pixel scale specifically comprises: for a target pixel and its eight neighbors, respectively extracting a generated branch input output by a generator and a real branch input extracted from a high-resolution remote sensing reflectance image, both of which are aligned one by one in terms of the number of wavebands and the position of the pixel; performing convolution degradation and down-sampling on each waveband of the generated branch input according to a pre-marked corresponding waveband point spread function, and then re-registering the generated branch input with the real branch input through bicubic interpolation to obtain a physically consistent comparison block; synchronously feeding the real branch input and the physically consistent comparison block into a convolution attention network of the same discriminator, retaining only a two-dimensional probability map output by a global average pooling layer, and indexing a center pixel real probability and a center pixel generated probability; calculating a cross-waveband first-order difference spectral gradient of the real branch input and the physically consistent comparison block, extracting a center pixel gradient modulus difference and combining a gradient direction cosine to obtain a multi-waveband gradient residual scalar; calculating a second-order Laplacian space gradient of the generated branch input and the real branch input, extracting a center pixel Laplacian value difference and comparing it with the neighborhood to obtain a Laplacian contrast residual, and normalizing the Laplacian contrast residual according to a spectral reflectance variance; determining a waveband confidence weight, performing a weighted average on the spectral gradient residual and a spatial coherence penalty, and combining the center pixel real generated probability difference to form a composite loss molecule; calculating a local water texture scale factor according to a visible light near-infrared water index difference of the target pixel and its neighbors, linearly normalizing the composite loss molecule, and outputting an adversarial consistency loss.

5. The remote sensing water quality pollution detection method based on super-resolution hybrid GAN of claim 4, wherein, From the high-resolution remote sensing reflectance image, a real spectral space block is extracted for a target pixel and its eight neighbors as a real branch input; at the same center pixel position, an inferred spectral space block of the corresponding size is synchronously extracted from a high-resolution residual signal output by a generator as a generated branch input; a center pixel explicit index is performed on a two-dimensional probability map to obtain a center pixel real probability and a center pixel generated probability; a center pixel Laplacian value difference is extracted and compared with the neighborhood four-connected centers to obtain a Laplacian contrast residual; and a waveband confidence weight is determined based on the reciprocal of the radiation signal-to-noise ratio of each waveband.

Images