Lake water quality parameter inversion method, medium, equipment and product based on remote sensing image

Abstract

The application discloses a lake water quality parameter inversion method based on remote sensing images, a medium, equipment and products, and relates to the field of water quality parameter inversion. The method comprises the following steps: acquiring Sentinel-2 remote sensing images and water quality sampling data, resampling bands in the remote sensing images to a unified resolution, and performing space-time matching on the resampled bands and the water quality sampling data; selecting double-band, triple-band and four-band combinations related to the water quality sampling data from the matched resampled bands, and constructing a training set; constructing a deep learning inversion model, which comprises three feature extraction subnetworks, an attention mechanism and a fusion main network; the band combinations are input into the three feature extraction subnetworks respectively, three feature vectors obtained through the three feature extraction subnetworks are input into the fusion main network after being weighted and fused through the attention mechanism, and a predicted water quality parameter concentration is output; the model is trained using the training set, and the trained model is used for lake water quality parameter inversion. The application improves the inversion accuracy of the water quality parameters.

Description

Technical Field

[0001] This invention relates to the field of water quality parameter inversion technology, and in particular to methods, media, equipment and products for inverting lake water quality parameters based on remote sensing images. Background Technology

[0002] Traditional water quality monitoring mainly relies on manual on-site sampling and laboratory analysis. Although it can obtain accurate data at a single point, it has obvious shortcomings: the monitoring space coverage is limited, making it difficult to fully reflect the spatial heterogeneity of lake water quality; the observation cycle is long and the update frequency is low, which cannot meet the needs of high-frequency dynamic monitoring; long-term large-scale sampling and testing are costly, and the sampling process itself may disturb the water body state, affecting the objectivity of the data.

[0003] Satellite remote sensing technology, with its advantages of wide coverage, high speed, non-contact operation, and low cost, has become an important means of achieving comprehensive water quality monitoring of lakes. The Sentinel-2 satellite, equipped with a multispectral imager, provides L2A-level data products with atmospheric correction, offering spatial resolutions of 10 meters, 20 meters, and 60 meters, and a short revisit period of approximately 5 days. This allows for the continuous acquisition of high-quality multispectral information, providing an excellent data foundation for water quality parameter retrieval.

[0004] However, current water quality inversion methods based on Sentinel-2 imagery still face certain limitations: in terms of feature construction, most rely on single bands or simple combinations, failing to fully explore the synergistic response information between multiple bands, resulting in insufficient spectral feature representation capabilities; in terms of model methods, traditional machine learning models (such as random forests, gradient boosting decision trees, etc.) have limited ability to fit the complex nonlinear relationship between water body spectra and water quality parameters, thus restricting further improvement in inversion accuracy. Summary of the Invention

[0005] The purpose of this invention is to address the problem that existing water quality inversion methods based on Sentinel-2 imagery rely on constructing features using a single band or a simple combination, failing to fully exploit the synergistic response information between multiple bands and resulting in insufficient spectral feature characterization capabilities. Therefore, this invention proposes a method for inverting lake water quality parameters based on remote sensing imagery, comprising the following steps:

[0006] S1. Acquire water quality sampling data and Sentinel-2 remote sensing images from each monitoring point of the target lake;

[0007] S2. Resample the bands of each resolution in the Sentinel-2 remote sensing image to a uniform resolution. Perform spatiotemporal matching of the surface reflectance data of the resampled bands with the water quality sampling data. Select the combination of two-band, three-band and four-band related to the water quality sampling data from the matched resampled bands to construct a water quality inversion training dataset.

[0008] S3. Construct a deep learning inversion model, including three feature extraction sub-networks, an attention mechanism, and a fusion main network. The dual-band, three-band, and four-band combinations are respectively input into the feature extraction sub-networks. The feature vectors output by the feature extraction sub-networks are weighted and fused by the attention mechanism. The weighted and fused feature vectors are input into the fusion main network, and the predicted water quality parameter concentrations are output.

[0009] S4. Use the water quality inversion training dataset to train the deep learning inversion model, and use the trained model to invert lake water quality parameters.

[0010] Furthermore, the water quality sampling data includes: water quality data, sampling time and sampling location. The water quality data includes: measured values ​​of chlorophyll-α, turbidity, chemical oxygen demand, ammonia nitrogen, total phosphorus and total nitrogen.

[0011] Furthermore, combinations of two-band, three-band, and four-band arrays relevant to the water quality sampling data were selected from the resampling bands, specifically:

[0012] Calculate the Pearson correlation coefficient between reflectance and water quality parameter concentration for each band combination, and select the top K combinations with the highest correlation to water quality parameter concentration from the two-band, three-band, and four-band combinations.

[0013] Furthermore, the two-band subnetwork is a neural network containing two fully connected layers, the three-band subnetwork is a neural network containing three fully connected layers, and the four-band subnetwork is a neural network containing three fully connected layers.

[0014] Furthermore, dual-band combination formats include:

[0015] , , , as well as ;

[0016] Three-band combination formats include:

[0017] and( )· ;

[0018] The four-band combination formats include:

[0019] , and ;

[0020] in, , , and These represent four different frequency bands.

[0021] The present invention also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0022] The present invention also proposes an electronic device, including a processor and a memory, wherein the processor is interconnected with the memory, the memory is used to store a computer program, the computer program including computer-readable instructions, and the processor is configured to invoke the computer-readable instructions to execute the above-described method.

[0023] The present invention also proposes a computer program product, including a computer program / instructions that, when executed by a processor, implement the above-described method.

[0024] The beneficial effects of the technical solution provided by this invention are:

[0025] (1) Highly correlated dual-band, three-band and four-band combination features were selected from Sentinel-2 multispectral data, which improved the spectral feature representation ability and provided richer and more powerful input for deep learning models.

[0026] (2) A multi-sub-network structure is adopted, and three different band combination forms are input into three sub-networks for feature extraction. Then, the network is aggregated through an attention mechanism to fully explore the collaborative response information between multiple bands, so that the model can learn the deep connection between different band combinations and significantly improve the inversion accuracy of water quality parameters. Attached Figure Description

[0027] Figure 1 This is a flowchart of a method for inverting lake water quality parameters based on remote sensing images, according to an example of the present invention.

[0028] Figure 2 This embodiment of the invention uses Python to visualize the raster matrix and employs a hierarchical coloring method to draw a two-dimensional spatial distribution map of turbidity and total phosphorus.

[0029] Figure 3 This is a classification of total phosphorus at different concentrations in embodiments of the present invention;

[0030] Figure 4 This is a block diagram of an electronic device according to an exemplary embodiment of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0032] A flowchart of a lake water quality parameter inversion method based on remote sensing imagery, as an example of the present invention, is shown below. Figure 1 Specifically, it includes:

[0033] S1. Acquire water quality sampling data and Sentinel-2 remote sensing images from each monitoring point of the target lake. Water quality sampling data includes: measured values ​​of water quality parameters (chlorophyll-α, turbidity, chemical oxygen demand, ammonia nitrogen, total phosphorus, total nitrogen, etc.) and their corresponding spatial locations (latitude and longitude coordinates) and sampling time. To ensure the reliability of the water quality inversion input data, firstly, from the available Sentinel-2 L2A image library, select images covering the target lake area that meet the following imaging quality conditions: the sky above the main body of the lake should be as cloudless or have few clouds as possible, and there should be no significant atmospheric scattering, haze, or sensor anomalies.

[0034] S2. Resample the bands of each resolution in the Sentinel-2 remote sensing image to a uniform resolution. Perform spatiotemporal matching of the surface reflectance data of the resampled bands with the water quality sampling data. Based on the matching results, select two-band, three-band, and four-band combinations related to water quality parameters. In this process, remove samples with extreme outliers in water quality concentration to optimize the quality and reliability of the dataset. Finally, construct a training dataset suitable for water quality inversion.

[0035] To fully utilize the multi-band information of Sentinel-2 and unify the spatial scale, the following processing is performed: using a resampling algorithm (such as bilinear interpolation), the bands with resolutions of 20 meters (B5, B6, B7, B8A, B11, B12) and 60 meters (B1, B9) in the Sentinel-2 L2A level image are uniformly upsampled to a spatial resolution of 10 meters, so that they are consistent with the spatial scale of high-resolution bands such as B2, B3, B4, and B8.

[0036] To ensure data comparability, the transit time of satellite imagery is strictly matched with the ground water quality sampling time, with the time difference typically controlled within ±8 hours to minimize errors introduced by changes in the water body itself. Each valid matching point constitutes a data sample. All the resampled bands are geometrically registered and synthesized to generate a full-band 10-meter resolution reflectance image with a unified geographic coordinate system (such as WGS84 UTM). In this embodiment of the invention, based on the precise geographic coordinates of the monitoring stations obtained in step 1, the surface reflectance values ​​of each station's corresponding pixel location in the eight core bands (B2, B3, B4, B5, B6, B7, B8, and B8A) are extracted from the synthesized image.

[0037] Remotely sensed reflectance data is correlated and paired with ground-measured water quality parameters to construct a one-to-one corresponding sample dataset. In water quality remote sensing inversion, the core purpose of band combination is to suppress the interference effects of atmosphere, illumination, sediment, and different water body components. Two-band combinations can amplify the complementary differences between band absorption and scattering, eliminate multiplicative interference from illumination, highlight single-group signal differences, accumulate target features, and strengthen relative response relationships. Three-band combinations can further separate the mixed spectral contributions of multiple water quality components, reflecting the interference suppression and feature separation relationships between bands. Four-band combinations are suitable for complex water body inversion scenarios, reflecting the joint comparison and multiple interference removal relationships between bands.

[0038] Based on the reflectance of the eight core bands of the Sentinel-2 imagery in the aforementioned water quality inversion dataset, two-band, three-band, and four-band combinations were generated, respectively. The methods for generating the two-band, three-band, and four-band combinations are as follows:

[0039] The five forms of dual-band combinations include:

[0040] , , , as well as ;

[0041] The two forms of three-band combination include:

[0042] and( )· ;

[0043] The three forms of four-band combinations include:

[0044] , and ;

[0045] in, , , and These represent four different bands out of the eight core bands.

[0046] Choosing two bands from eight core bands yields 28 options, which are then combined in five different dual-band combinations. Choosing three bands from eight core bands yields 56 options, which are then combined in two different three-band combinations. Choosing four bands from eight core bands yields 70 options, which are then combined in three different four-band combinations. The sample of this invention includes: 140 dual-band combinations, 336 three-band combinations, and 2296 four-band combinations.

[0047] For target water quality parameters (such as turbidity or total phosphorus), the Pearson correlation coefficient between reflectance and water quality parameter concentration values ​​for each band combination is calculated to quantify the indicative ability of each band combination to indicate changes in water quality parameters. The calculation formula is as follows:

[0048]

[0049] in Let i be the reflectance value of a certain band combination for the i-th sample. Let n represent the concentration values ​​of the corresponding water quality parameters, and n be the total number of samples. From the two-band, three-band, and four-band combinations, the top 10 combinations with the highest correlation to the water quality parameters were selected. During this process, null values ​​or extreme outliers in the water quality parameter samples were removed to improve the overall quality and reliability of the dataset. Finally, a water quality dataset containing 30 highly correlated band combinations was constructed. This dataset will serve as the input data source for the subsequent water quality parameter inversion model, with the model input features using the selected 30 highly correlated band combinations.

[0050] S3. Construct a deep learning inversion model, including three feature extraction sub-networks, an attention mechanism, and a fusion main network. The dual-band, three-band, and four-band combinations are respectively input into the feature extraction sub-networks. The feature vectors output by the feature extraction sub-networks are weighted and fused by the attention mechanism. The weighted and fused feature vectors are input into the fusion main network, which outputs the predicted water quality parameter concentrations.

[0051] The network consists of three subnetworks: a two-band subnetwork (128-dimensional, 64-dimensional hidden layer), a three-band subnetwork (256-dimensional, 128-dimensional, 64-dimensional hidden layer), and a four-band subnetwork (512-dimensional, 256-dimensional, 128-dimensional hidden layer). Each subnetwork uses ReLU activation, batch normalization, and Dropout (0.2% Dropout rate) after its hidden layers, outputting a 10-dimensional feature vector. The 30-dimensional feature vectors from the three subnetworks are weighted and fused using a modal-level attention mechanism. This attention mechanism learns three weights from the 30-dimensional vector using a linear layer and a softmax function; these weights represent the overall importance of each subnetwork's output. Each weight is then expanded to 10 dimensions (matching the output dimension of the corresponding subnetwork) and multiplied element-wise with the corresponding subnetwork's 10-dimensional output to obtain the weighted feature vector. Finally, the three weighted 10-dimensional features are concatenated into a final 30-dimensional fused feature vector. This mechanism assigns dynamic weights to the modal features output by each sub-network to highlight feature combinations that contribute more to the inversion task. The weighted feature vector is input into the fusion main network, which contains three fully connected layers (512-dimensional, 256-dimensional, and 128-dimensional hidden layers). Through multi-layer nonlinear transformations, information from different dimensional band combinations is fused, and a scalar value is finally output, which is the predicted water quality parameter concentration.

[0052] S4. Use the water quality inversion training dataset to train the deep learning inversion model, and use the trained model to invert lake water quality parameters.

[0053] Before inputting the Sentinel-2 remote sensing image into the model, the Otsu automatic thresholding method was used to perform preliminary water body extraction based on the near-infrared band (B8) of the image. After converting the preliminary water body extraction results into SHP vector format, the precise water body vector specific to the lake was obtained by spatial overlay and cropping in ArcMap. Using this water body vector as a mask, the preprocessed Sentinel-2 full-band image was cropped to finally obtain a standardized model input image containing only pure lake water.

[0054] Pure water body imagery is input into a pre-trained deep learning inversion model to execute the inversion process. Based on a band combination strategy, the model reads the reflectance values ​​of each water body pixel in eight core bands and organizes them into corresponding feature combinations for input. The model calculates the predicted concentration value of the corresponding water quality parameter for each water body pixel.

[0055] The process iterates through all water body pixels, arranging the predicted concentration values ​​according to their original geographic coordinates to generate a two-dimensional concentration raster matrix that perfectly matches the geographic extent, spatial resolution (10 meters), and coordinate system of the input image. Based on the statistical distribution characteristics of the concentration values, an appropriate classification method (such as the natural breakpoint method) is selected for grading, and each grade is assigned a visually distinctive color to form a thematic color map. Finally, the output is a high-resolution image file or a raster file with geographic information, thus obtaining a high-resolution spatial distribution map of the target lake's water quality parameter concentrations.

[0056] To verify the feasibility of this invention, Sentinel-2 L2A level remote sensing image data of the study area from February 2021 to December 2024 were collected. To ensure matching with ground-measured data, the screening criteria were set as follows: at least one monitoring station in the study area was not covered by clouds, cloud shadows, or dense fog, and effective water reflectance could be extracted. Based on this criterion, a total of 104 valid images were acquired. The L2A level image data were loaded using SNAP software, and the Resampling tool was called. A bilinear interpolation algorithm was used to upsample the 20-meter and 60-meter resolution bands to 10 meters, and the data was exported in ENVI format.

[0057] All 10-meter resampling bands were loaded using ENVI software, and a full-band 10-meter resolution image was synthesized using the Build layer Stack tool, with the coordinate system WGS84-UTM50N. Reflectance extraction: Based on the latitude and longitude coordinates of the four monitoring stations, the reflectance values ​​of the eight core bands (B2, B3, B4, B5, B6, B7, B8, and B8A) for the corresponding pixels of each station were extracted using the SpectralProfile tool in ENVI software. The reflectance data was matched with water quality data according to sampling time, with the time window controlled within ±8 hours, to form an initial dataset, yielding 383 valid samples (some samples had missing individual water quality parameters).

[0058] Turbidity (an optical parameter) and total phosphorus (a non-optical parameter) were selected as targets for subsequent water quality parameter demonstrations. Using a band combination method, two-band, three-band, and four-band combinations were generated. The top 10 combinations with the highest absolute values ​​of Pearson correlation coefficients with the target water quality parameters were selected, and the calculation formulas for these combinations were fixed and saved. During this process, missing and outlier values ​​were removed from the data for both water quality parameters, ultimately yielding 353 sets of turbidity data and 362 sets of total phosphorus data. Therefore, a total of 30 highly correlated band combination features were selected for each water quality parameter as input for subsequent deep learning models.

[0059] After the model training of this invention is completed, to verify the advantages of the multi-network fusion architecture of this invention, the model is compared with several typical inversion methods, including Random Forest, XGBoost, CATBoost, and Gradient Boosting Decision Tree machine learning models. All comparison methods use the same 30 sets of band combination features selected by the process of this invention as input to ensure fairness in the starting point of features. The three feature extraction sub-networks of this invention perform differentiated feature learning on these 30 sets of band combinations: the first sub-network focuses on dual-band combinations with strong linear responses, the second sub-network models the nonlinear coupling relationship between three bands, and the third sub-network captures high-order spectral interaction features in four-band combinations. After the three features are weighted and fused through an attention mechanism, they are input into a shared fully connected layer for further feature extraction and inversion, achieving multi-scale and multi-level feature complementarity and significantly improving the accuracy of turbidity and total phosphorus inversion. The comparison model directly extracts overall features from the 30 sets of band combinations, without demonstrating differentiated feature learning.

[0060] As shown in Table 1, the method of this invention exhibits the best accuracy in turbidity inversion; as shown in Table 2, the method of this invention also performs best in total phosphorus inversion. Experimental results show that, compared with traditional methods, the model of this invention has significant improvements in evaluation indicators such as coefficient of determination, root mean square error, and mean absolute error, verifying the advantages of the multi-network fusion architecture in mining synergistic information between different types of band combination features.

[0061] Table 1

[0062]

[0063] Table 2

[0064]

[0065] Based on the trained deep learning inversion model and the remote sensing image data to be inverted, a two-dimensional spatial distribution map of the target lake's water quality parameters was generated. From 104 acquired images, 67 high-quality images with little or no cloud cover were selected over the study area. A Sentinel-2 remote sensing image from September 1, 2023, was chosen as the data source. Before inputting the image into the model, preliminary water body extraction was performed using the Otsu automatic thresholding method based on the near-infrared B8 band. The preliminary water body extraction results were converted into SHP vector format and refined to obtain a precise water body vector specific to the lake. Using this precise water body vector as a mask, the preprocessed Sentinel-2 full-band image was cropped to finally obtain a standardized model input image containing only the pure lake water area. This pure water area image was input into the trained deep learning water quality inversion model for pixel-by-pixel forward inference, generating raster data of the spatial distribution of water quality parameter concentrations with a spatial resolution of 10 meters.

[0066] Two-dimensional spatial distribution map generation: Using Python to visualize the raster matrix, a hierarchical color scheme is used to draw a two-dimensional spatial distribution map of turbidity and total phosphorus. For example... Figure 2 As shown, this figure illustrates the turbidity classification at different concentrations, with pixels greater than 160 NTU uniformly represented by a single color. The turbidity inversion map of the study area clearly displays the spatial distribution characteristics of turbidity in the lake region; as shown... Figure 3 As shown, the figure displays the total phosphorus classification at different concentrations, with pixels greater than 90 mg / m³ represented by a single color. The total phosphorus inversion map of the study area clearly shows the spatial distribution characteristics of total phosphorus in the lake area.

[0067] In one exemplary embodiment, a computer-readable storage medium is included, which stores a computer program that, when executed by a processor, implements the method described above.

[0068] Please see Figure 4 In one exemplary embodiment, the device further includes an electronic device including at least one processor, at least one memory, and at least one communication bus.

[0069] The memory stores a computer program, which includes computer-readable instructions. The processor calls the computer-readable instructions stored in the memory through a communication bus to execute the above method.

[0070] In one exemplary embodiment, a computer program product is proposed, including a computer program / instructions that, when executed by a processor, implement the method described above.

[0071] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for inverting lake water quality parameters based on remote sensing images, characterized in that, Includes the following steps: S1. Acquire water quality sampling data and Sentinel-2 remote sensing images from each monitoring point of the target lake; S2. Resample the bands of each resolution in the Sentinel-2 remote sensing image to a uniform resolution. Perform spatiotemporal matching of the surface reflectance data of the resampled bands with the water quality sampling data. Select two-band combinations, three-band combinations and four-band combinations related to the water quality sampling data from the matched resampled bands to construct a water quality inversion training dataset. S3. Construct a deep learning inversion model, including three feature extraction sub-networks, an attention mechanism, and a fusion main network. The two-band combination, three-band combination, and four-band combination are respectively input into the three feature extraction sub-networks. The feature vectors output by the feature extraction sub-networks are weighted and fused by the attention mechanism. The weighted and fused feature vectors are input into the fusion main network, and the predicted water quality parameter concentrations are output. S4. Use the water quality inversion training dataset to train the deep learning inversion model, and use the trained model to invert lake water quality parameters.

2. The method for inverting lake water quality parameters based on remote sensing images according to claim 1, characterized in that, Water quality sampling data includes: water quality data, sampling time and sampling location. The water quality data includes: measured values ​​of chlorophyll-α, turbidity, chemical oxygen demand, ammonia nitrogen, total phosphorus and total nitrogen.

3. The method for inverting lake water quality parameters based on remote sensing images according to claim 1, characterized in that, From the resampling bands, select combinations of two-band, three-band, and four-band bands that are relevant to the water quality sampling data, specifically: Calculate the Pearson correlation coefficient between reflectance and water quality parameter concentration for each band combination, and select the top K combinations with the highest correlation to water quality parameter concentration from the two-band, three-band, and four-band combinations.

4. The method for inverting lake water quality parameters based on remote sensing images according to claim 1, characterized in that, A two-band subnetwork is a neural network containing two fully connected layers, a three-band subnetwork is a neural network containing three fully connected layers, and a four-band subnetwork is a neural network containing three fully connected layers.

5. The method for inverting lake water quality parameters based on remote sensing images according to claim 1, characterized in that, Dual-band combination formats include: , , , as well as ; Three-band combination formats include: and( )· ; The four-band combination formats include: , and ; in, , , and These represent four different frequency bands.

6. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 5.

7. An electronic device, characterized in that, The device includes a processor and a memory, the processor being interconnected with the memory, wherein the memory is used to store a computer program, the computer program including computer-readable instructions, and the processor is configured to invoke the computer-readable instructions to perform the method as described in any one of claims 1 to 5.

8. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the method described in any one of claims 1 to 5.

Images