High / multi-spectral collaborative chlorophyll-a retrieval method based on feature-level residual learning

Abstract

The application discloses a kind of high / multi-spectral collaborative chlorophyll-a inversion methods based on feature level residual learning.Firstly, the application utilizes multispectral wide band feature and measured chlorophyll-a true value to train regularization regression model, and the true value and predicted value are subtracted, and residual error vector is constructed;Subsequently, using high spectral high frequency derivative feature, multivariate dimension reduction regression model is trained with this residual as target;The multispectral feature matrix of each pixel of the target water body is input into the trained regularization regression model, and the baseline predicted value of the chlorophyll-a concentration of each pixel is obtained;The high spectral derivative feature matrix of each pixel is input into the trained multivariate dimension reduction regression model, and the high frequency residual correction value of each pixel is obtained;Finally, the two values are added, and the final predicted value of chlorophyll-a concentration is obtained after hyperbolic tangent soft threshold function post-processing, and the inversion graph is output.The application realizes the bottom breakthrough of multi-source fusion paradigm, has strong physical noise resistance, and improves the high concentration chlorophyll-a inversion precision.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing technology of chlorophyll a concentration in water bodies, specifically involving a hyperspectral / multispectral synergistic chlorophyll a inversion method based on feature-level residual learning. Background Technology

[0002] Accurate monitoring of chlorophyll a (Chl-a) concentrations in water bodies is crucial for global water security and biogeochemical cycles. Over the past few decades, satellite remote sensing has become an indispensable tool for large-scale monitoring of chlorophyll a in complex water bodies. The core bio-optical mechanism of chlorophyll a retrieval is highly dependent on the relative contrast between the maximum pigment absorption in the red band (approximately 665 nm) and the prominent fluorescence and particle backscattering peaks in the red-edge to near-infrared region (700-750 nm).

[0003] Currently, broadband multispectral sensors (such as the MSI of Sentinel-2) have become the mainstay of inland water monitoring due to their high spatial resolution. Meanwhile, emerging spatial hyperspectral imagers (such as the GF-5 AHSI) can provide continuous narrow bands, perfectly resolving high-frequency bio-optical curvature and accurately locating fluorescence peaks. To overcome the limitations of single sensors, multi-source sensor fusion (combining the spatial / radiative stability of multispectral sensors with the spectral sensitivity of hyperspectral sensors) has become a cutting-edge technology in remote sensing.

[0004] Despite the theoretical potential of multi-source fusion, existing multi-source fusion techniques (such as "radiative cross-calibration" and "mathematical virtual band reconstruction") suffer from the following insurmountable physical pitfalls and drawbacks when applied to optically complex eutrophic water bodies: 1. The "spectral averaging effect" of multispectral sensors leads to signal saturation: As eutrophication intensifies, the strong backscattering from phytoplankton combined with absorption in pure water triggers a significant "redshift" effect, pushing the optimal diagnostic reflectance peak to a longer wavelength (approximately 717 nm). Existing multispectral sensors (such as Sentinel-2) have excessively broad bands in the red-edge region (e.g., Band 5 bandwidth of 15 nm), inevitably integrating and smoothing the originally sharp fluorescence peaks (i.e., the "spectral averaging effect"). This spectral truncation weakens the sensitivity of classical band ratio algorithms, leading to severe signal saturation and a systematic underestimation of chlorophyll a concentration in highly eutrophic regions.

[0005] 2. Traditional radiation alignment strategies disrupt local spectral curvature (radiative homogenization trap): To address the issue of radiometric inconsistencies between different sensors, existing pixel-level radiometric alignment techniques (such as empirical linear calibration or PCHIP interpolation) force hyperspectral data to match multispectral baselines on an absolute radiometric scale. This operation incorrectly suppresses and flattens genuine high-frequency fluorescence peaks in the hyperspectral data, completely destroying the "local shape consistency" upon which semi-analytical inversion models heavily rely. Since band ratio algorithms rely on the relative contrast between adjacent narrowbands to eliminate background noise, arbitrarily scaling the absolute radiance will erase the high-frequency physical signals required for inversion.

[0006] 3. Mathematical virtual band reconstruction causes a severe "noise amplification effect": Another existing technique involves using linear combinations of broad multispectral bands to mathematically synthesize virtual narrow hyperspectral bands (such as a virtual 717 nm). However, according to error propagation theory, actual satellite observation data inevitably contains additive atmospheric residual errors. Linear combination operations indiscriminately accumulate and amplify these additive atmospheric noises, leading to a sharp drop in the signal-to-noise ratio of the virtual features. Under complex actual satellite transmission conditions, this "soft enhancement" method, which relies on purely mathematical means, often exhibits performance inversion and is even less stable than the native multispectral ratio model.

[0007] In summary, existing technologies cannot effectively decouple and integrate the radiative spatial stability of multispectral and high-frequency spectral fidelity of hyperspectral without distorting spectral features and amplifying additive noise, resulting in generally low accuracy of chlorophyll a spatial gradient inversion in complex inland water bodies. Summary of the Invention

[0008] To address the issues of existing technologies, such as the "spectral averaging effect" in multispectral sensors leading to saturation of high-concentration chlorophyll-a (Chl-a) inversion signals in complex water bodies, and the susceptibility of traditional multi-source remote sensing fusion methods to "radiative homogenization traps (destroying local spectral curvature)" and "noise amplification traps (indiscriminate accumulation of atmospheric additive noise)" during radiometric cross-calibration and virtual band reconstruction, this invention provides a hyperspectral / multispectral synergistic chlorophyll-a inversion method based on feature-level residual learning. Unlike traditional radiometric cross-calibration algorithms, this invention, through multispectral macroscopic baseline anchoring and hyperspectral high-frequency residual prediction, abandons the traditional approach of forcibly unifying the radiometric scale. It decouples the inversion process into macroscopic spatial baseline anchoring and high-frequency spectral residual refinement, accurately restoring the true biogeochemical spatial gradient of eutrophic water bodies while being immune to atmospheric noise and absolute radiometric instability.

[0009] This invention acquires multi-source satellite imagery and measured water quality data of a target water body, trains a regularized regression model using multispectral broadband data, calculates the difference between the measured concentration and the baseline predicted value to obtain a residual error vector, uses the residual error vector as the target variable, and trains a multivariate dimensionality reduction regression model using hyperspectral high-frequency derivative features, acquires panoramic multispectral and hyperspectral remote sensing images of unknown concentrations in the target water body, inputs the multispectral feature matrix of each pixel into the trained regularized regression model to obtain the baseline predicted value of chlorophyll a concentration for each pixel, inputs the hyperspectral derivative feature matrix of each pixel into the trained multivariate dimensionality reduction regression model to obtain the high-frequency residual correction value for each pixel, and uses the sum of the baseline predicted value of chlorophyll a concentration and the high-frequency residual correction value of each pixel as the initial co-predicted value, performs soft threshold function constraint processing on it, and obtains the final predicted value of chlorophyll a concentration for each pixel.

[0010] The specific steps are as follows: Step 1: Acquire multi-source satellite imagery and measured water quality data, perform image preprocessing and spatiotemporal consistency matching, and construct a dataset for model training and validation.

[0011] Furthermore, when acquiring multi-source satellite imagery and measured water quality data for the target water body, atmospheric correction and radiometric calibration are performed on the acquired multi-source satellite imagery (including multispectral and hyperspectral imagery) to obtain remote sensing reflectance data. Using the multispectral imagery as a spatial reference, the hyperspectral imagery is spatially resampled and geometrically registered to ensure consistent spatial resolution between the two.

[0012] Furthermore, in order to avoid the pollution of near-infrared signals by the "land proximity effect" of the land-water boundary of inland lakes when acquiring multi-source satellite images and measured water quality data of target water bodies, the hyperspectral images are thresholded using an improved normalized difference water body index to generate a hyperspectral initial water body mask, and the multispectral images are generated by combining red band reflectance threshold with connected component analysis.

[0013] Preferably, when acquiring multi-source satellite images and measured water quality data of the target water body, an inward buffering corrosion operation of 1 pixel is performed on the initial water body mask to extract pure water pixels located within the corrosion mask, thereby improving the purity of the water body spectrum.

[0014] Furthermore, when acquiring multi-source satellite imagery and measured water quality data of the target water body, in order to eliminate the damage to subsequent derivative feature extraction caused by the inherent stripe noise and high-frequency jitter of hyperspectral images, a continuous filtering fidelity strategy is adopted. Spatial median filtering and spectral smoothing filtering (such as Savitzky-Golay filtering) are continuously applied to the extracted hyperspectral remote sensing reflectance data cube to preserve the intrinsic spectral curvature.

[0015] Step 2: Train a regularized regression model using multispectral broadband data: Extract broadband remote sensing reflectance from the multispectral image data in the training set and calculate the normalized differential chlorophyll index (NDCI); concatenate the broadband reflectance and the calculated NDCI into a multispectral feature matrix; use the multispectral feature matrix as the input variable and the corresponding measured chlorophyll a concentration data as the target variable to train the ridge regression model; after training, input the multispectral feature matrix into the trained regularized regression model to output the baseline predicted value of chlorophyll a concentration; calculate the difference between the measured concentration and the baseline predicted value to obtain the residual error vector; As a preferred method, the multispectral feature matrix is ​​standardized before model training; As a preferred option, during model training, to penalize the large coefficients caused by wide-band multicollinearity, an L2 regularization parameter (such as α=0.1) is set for fitting. The high signal-to-noise ratio of multispectral data is used to construct a "macro baseline" that is spatially stable and resistant to multiplicative noise.

[0016] Step 3: Train the multivariate dimensionality reduction regression model using hyperspectral high-frequency derivative features: Extract a continuous narrow-band matrix from the red-edge region (wavelength range of approximately 650-750 nm) of the hyperspectral image data in the training set; calculate the first-order geometric gradient features (used to eliminate atmospheric additive bias) and the second-order geometric gradient features (used to eliminate linear background scattering) of this matrix along the spectral dimension, and concatenate the first-order and second-order derivative features in the column direction to form a hyperspectral derivative feature matrix; use the hyperspectral derivative feature matrix as the input variable and the residual error vector from Step 2 as the target variable to train the multivariate dimensionality reduction regression model.

[0017] Without any forced alignment of radiation scales or the use of purely mathematical methods to create virtual bands, this method leverages the high-frequency geometric curvature properties of the hyperspectral derivative, making it naturally immune to atmospheric additive bias and absolute radiation errors. Through a multivariate dimensionality reduction regression model, it uses pure high-frequency signals to accurately diagnose and predict the residual errors in the regularized regression model caused by the "spectral averaging effect," outputting high-frequency residual correction values.

[0018] Step 4: High-precision inversion of chlorophyll a concentration from panoramic remote sensing images of the target water body: Acquire panoramic multispectral and hyperspectral remote sensing images of the target water body, preprocess the acquired images and perform spatiotemporal consistency matching; traverse all valid pure water body pixels, input the multispectral feature matrix of each pixel into the regularized regression model trained in Step 2 to obtain the baseline predicted value of chlorophyll a concentration for each pixel; input the hyperspectral derivative feature matrix of each pixel into the multivariate dimensionality reduction regression model trained in Step 3 to obtain the high-frequency residual correction value for each pixel; use the sum of the baseline predicted value of chlorophyll a concentration and the high-frequency residual correction value of each pixel as the initial collaborative prediction value, and then perform soft threshold function constraint processing on it to obtain the final predicted value of chlorophyll a concentration for each pixel, and output a high-precision collaborative chlorophyll a inversion map of the panoramic view of the target water body.

[0019] By using the dual-stream collaborative framework, not only is the macroscopic spatial texture distribution of the original multispectral image preserved, but the systematic underestimation defect of the single baseline model in high-concentration areas before optimization is also more effectively corrected.

[0020] Furthermore, before making a prediction, step four involves performing spatiotemporal registration, pure water mask extraction, and filtering and denoising on the acquired panoramic multispectral and hyperspectral remote sensing images of the target water body.

[0021] Furthermore, the process of performing soft threshold function constraint processing in step four is as follows: Set the linear threshold and physical limit extreme value of the target water body, take the sum of the baseline predicted value of chlorophyll a concentration of each pixel and the high-frequency residual correction value as the initial co-predicted value, remove the negative values ​​in the initial co-predicted value, and then perform soft threshold function constraint processing on it to smoothly compress the initial predicted value that is greater than the linear threshold and map it to the range of physical limit extreme value, so as to obtain the final predicted value of chlorophyll a concentration of each pixel, and output a high-precision co-chlorophyll a inversion map of the target water body panorama.

[0022] This invention achieves a fundamental breakthrough in the multi-source fusion paradigm (solving the radiation homogenization trap). Breaking away from the traditional mindset of forcing hyperspectral to multispectral "radiative alignment," it instead adopts a "feature-level mechanism complementarity." By allowing multispectral to "dominate" (responsible for macroscopic spatial baseline and signal-to-noise ratio stability) and hyperspectral to "dominate" (capturing localized minute curvatures smoothed out in the red-edge region through derivative features to compensate for saturation residuals), the spatial stability and spectral sensitivity tasks are successfully decoupled.

[0023] The method described in this invention has extremely strong physical noise immunity (solving the noise amplification trap). By using the first and second derivatives of the hyperspectral red edge region as input features, it is naturally immune to constant atmospheric additive bias, linear background scattering, and the absolute radiation instability of the sensor itself at the mathematical and physical level; it effectively avoids the fatal flaw of traditional virtual band linear reconstruction that indiscriminately amplifies additive atmospheric noise, leading to model collapse.

[0024] The method described in this invention has significantly improved the accuracy of high-concentration chlorophyll retrieval. To address the redshift of the fluorescence peak at 717 nm in eutrophic lakes and the spectral averaging effect (signal smoothing loss) caused by the multispectral broadband, this invention compensates for the residual upward pull, accurately diagnosing and correcting the systematic underestimation of the multispectral spectrum. Attached Figure Description

[0025] Figure 1 This is a flowchart of the method described in this invention; Figure 2 This is a diagram showing the correction and optimization effect of the dual-stream collaborative framework in the embodiment; Figure 3 The output of this example is a spatial distribution inversion map of chlorophyll a concentration in the target water body and a comparison map of its latitude average profile with that of a single multispectral baseline model. Figure 4 This is a comparison chart showing the accuracy evaluation of the method described in this invention with that of a single sensor and a traditional fusion strategy in the embodiments. Detailed Implementation

[0026] This embodiment uses Erhai Lake in China as a specific application scenario, and combines the actual satellite operating environment and numerical parameters to demonstrate and explain the technical solution of the present invention in detail. A hyperspectral / multispectral synergistic chlorophyll a inversion method based on feature-level residual learning specifically includes the following steps: Step 1: Acquisition and preprocessing of multi-source satellite data (constructing a model training dataset) (1) Data Acquisition: Sentinel-2 MSI (multispectral) and GF-5 AHSI (hyperspectral) Level-1 images of Erhai Lake with a time window difference of ±1 day were acquired, along with synchronously collected measured water quality data (a total of N = 43 valid sampling points were matched, and the measured chlorophyll a concentration ranged from 14.6 to 86.9 mg m³). -3 ).

[0027] (2) Atmospheric correction and registration: Sentinel-2 L1C data were atmospherically corrected using ACOLITE software (DSF algorithm), and GF-5 L1 data were atmospherically corrected using the FLAASH model. The spatial resolution of Sentinel-2 was resampled to 30m and Auto-GCP geometric co-registration was performed with GF-5.

[0028] (3) Water masking and anti-proximity effect processing: Two initial masks were generated. For Sentinel-2 images, a red band reflectance threshold B4 > 1500 was set and connected component analysis was used to extract multispectral initial water bodies. For GF-5 images, a modified Normalized Difference Water Index (MNDWI) was used for threshold segmentation to extract hyperspectral initial water bodies. Subsequently, an inward buffer erosion operation of 1 pixel (30 m) was forced on both initial masks. After finding the intersection, only deep pure water body pixels were retained.

[0029] (4) Hyperspectral signal fidelity and filtering: After extracting the GF-5 remote sensing reflectance, 3×3 spatial median filtering and Savitzky-Golay (SG) spectral filtering were applied in sequence (sliding window length = 9, polynomial fitting order = 2) to remove high-frequency sensor jitter noise while accurately preserving the intrinsic spectral curvature of the red edge region.

[0030] (5) Extracting the training set: Based on the latitude and longitude coordinates of the 43 measured sampling points, the corresponding multispectral and hyperspectral pixel reflectances are extracted, and a multispectral feature matrix, a hyperspectral feature matrix and a measured concentration vector with a dimension of 43 are constructed respectively, which are used as training data for the subsequent model.

[0031] Step 2: Train a regularized regression model using multispectral broadband data (in this example, the Ridge Regression model is used as an example): From the 43-dimensional multispectral feature matrix, the reflectance matrices X of the red B4 (665nm), red-edge B5 (705nm), and near-infrared B6 (740nm) bands of Sentinel-2 are extracted. s2_feats .

[0032] Calculate the Normalized Differential Chlorophyll Index (NDCI): NDCI = (B5 - B4) / (B5 + B4), and include it as the fourth characteristic in X. s2_feats .

[0033] For X s2_feats After standardization, the 43 measured chlorophyll a concentration vectors were used as input variables and as target variables to train the Ridge Regression model. During the training process, the L2 regularization penalty coefficient α = 0.1 was set to suppress collinearity across wide bands.

[0034] Standardized X s2_feats The input to the trained Ridge Regression model is the vector Y, which represents the baseline predicted chlorophyll a concentration from 43 sampling points. baseThen, the difference between the measured concentration and the baseline predicted value is calculated to obtain the residual error vector: This is used as the target for training the Partial Least Squares Regression (PLSR) model.

[0035] Step 3: Train a multivariate dimensionality reduction regression model using hyperspectral high-frequency derivative features (in this example, the partial least squares regression (PLSR) model is used as an example). Define the range of the hyperspectral red edge region: Extract a continuous narrow band with indices from 60 to 90 (corresponding to a wavelength range of approximately 650-750 nm) from the GF-5 hyperspectral feature matrix of dimension 43, denoted as matrix X. gf5_roi .

[0036] The first derivative D1 (to eliminate atmospheric additive bias) and the second derivative D2 (to eliminate linear scattering tendency) of this matrix are calculated along the spectral dimension, and then the two are concatenated column-wise to form a hybrid derivative shape characteristic matrix X. gf5_shape = [D1, D2].

[0037] X above gf5_shape The matrix is ​​used as the input variable, with the residual error vector obtained in step two. The partial least squares regression (PLSR) model was trained using the target variable for prediction.

[0038] The number of potential principal components (n_components) was set to 3, and the PLSR model was fitted and trained. After training, the optimized model weights and parameters were obtained and saved to construct the final high-frequency residual correction model for subsequent pixel-level prediction of panoramic images.

[0039] Step 4: High-precision inversion of chlorophyll a concentration from panoramic remote sensing images of the target water body: Sentinel-2 and GF-5 remote sensing images of the entire Erhai Lake area were acquired, and the water body masking process in step one was performed on the remote sensing images to obtain the reflectance data of all effective water body pixels in the entire lake.

[0040] Iterate through all valid pure water pixels, input the multispectral feature matrix of each pixel into the Ridge Regression model trained in step two, and obtain the baseline predicted value Y of chlorophyll a concentration for each pixel. base_map Input the hyperspectral derivative feature matrix of each pixel into the partial least squares regression (PLSR) model trained in step three to obtain the high-frequency residual correction value Y for each pixel. resid_map ; A reliable linear threshold T = 60 mg m is set for the Erhai Lake area. -3 The physical limit is M = 100 mg m -3The baseline predicted value Y of chlorophyll a concentration for each pixel. base_map With high-frequency residual correction value Y resid_map The sum is used as the initial collaborative prediction value. Negative values ​​in the initial collaborative prediction value are removed (if less than 0, they are truncated to 0). Then, a hyperbolic tangent soft thresholding function is applied to it for post-processing: for the initial prediction value Y raw For data points > 60, apply formula Y final = 60 + (100-60)×tanh[(Yraw-60) / (100-60)] for smooth compression; for Y raw Data points ≤ 60 remain unchanged. The final predicted value of chlorophyll a concentration for each pixel is obtained, and a high-precision co-chromatic chlorophyll a inversion map of the target water body is output.

[0041] The optimization effect of the above dual-stream collaborative correction is as follows: Figure 2 As shown, Figure 2 The middle (a) is the inversion correction vector diagram from the multispectral baseline to the final synergistic result, showing the inversion improvement from the Sentinel-2 single baseline (gray circle) to the final synergistic result (red triangle). The vertical gray line represents the magnitude and direction of the residual correction for the GF-5 derivative feature contribution. Figure 2 (b) is a scatter plot showing the relationship between the true residuals of the multispectral model and the hyperspectral prediction corrections, illustrating the relationship between the baseline true residuals and the GF-5 prediction corrections. Figure 2 It can be seen that the red-edge derivative feature of GF-5 successfully captures the systematic error of the multispectral baseline, especially in the high-concentration region (first quadrant). The residual correction term, as an upward vertical vector, effectively pulls the originally severely underestimated prediction value back to the 1:1 ideal inversion line.

[0042] like Figure 3 As shown, Figure 3 (a) is the whole lake chlorophyll a inversion map based solely on the Sentinel-2 broadband band. As can be seen from the figure, the baseline inversion map based solely on the Sentinel-2 broadband band significantly underestimates the chlorophyll a concentration in the highly eutrophic northern bay due to signal saturation, thus losing spatial details. Figure 3 (b) shows the spatial residual correction (Δ_GF5) predicted by the high-frequency residual correction model. The red area intuitively reveals that the hyperspectral derivative features of this invention successfully provide significant and accurate positive concentration boosting and compensation for the northern algal bloom hotspot area, which was severely underestimated by the multispectral baseline due to signal saturation before optimization. Figure 3 Image (c) is the final high-precision inversion image generated after applying the method described in this invention in the embodiment, combined with... Figure 3The latitudinal average profile (d) shows that the method described in this invention successfully decoupled spatial texture from spectral sensitivity, preserving not only the macroscopic spatial gradient of the entire lake but also successfully identifying and accurately reconstructing the strong Microcystis bloom hotspots (>100 mg m) in the northern bay that were previously obscured by multispectral signal saturation. -3 This achieves truly high-precision remote sensing mapping.

[0043] Furthermore, to demonstrate the optimization level of the method described in this invention compared to traditional methods, it was compared with the native multispectral model and six existing fusion inversion strategies. For example... Figure 4 As shown, Figure 4 (a) shows the density scatter plot and frequency distribution histogram of the dual-flow cooperative model; Figure 4 In (b), the coefficient of determination R represents the different inversion strategies. 2 A bar chart comparing the root mean square error (RMSE). Figure 4 It can be seen that before optimization, using a single Sentinel-2 multispectral native baseline model (or a traditional virtual band reconstruction algorithm), due to the "spectral averaging effect" and "noise amplification effect," the inversion accuracy (Ri) is low. 2 The R value remained only around 0.41 to 0.55; however, after adopting the dual-stream cooperative inversion model described in this invention, R... 2 A significant leap was achieved, exceeding 0.75, while the root mean square error (RMSE) was significantly reduced to 6.66 mg m. -3 It comes very close to the inversion limit in in-situ hyperspectral theory.

Claims

1. A hyperspectral / multispectral synergistic chlorophyll a inversion method based on feature-level residual learning, characterized in that, Specifically as follows: Multi-source satellite imagery and measured water quality data of the target water body were acquired. A regularized regression model was trained using multispectral broadband data. The difference between the measured concentration and the baseline predicted value was calculated to obtain the residual error vector. The residual error vector was used as the target variable, and a multivariate dimensionality reduction regression model was trained using hyperspectral high-frequency derivative features. A panoramic multispectral and hyperspectral remote sensing image of the target water body with unknown concentrations was acquired. The multispectral feature matrix of each pixel was input into a trained regularized regression model to obtain the baseline predicted value of chlorophyll a concentration for each pixel. The hyperspectral derivative feature matrix of each pixel was input into a trained multivariate dimensionality reduction regression model to obtain the high-frequency residual correction value for each pixel. The sum of the baseline predicted value of chlorophyll a concentration and the high-frequency residual correction value of each pixel was used as the initial co-predicted value. A soft threshold function constraint was applied to it to obtain the final predicted value of chlorophyll a concentration for each pixel.

2. The hyperspectral / multispectral synergistic chlorophyll a inversion method based on feature-level residual learning as described in claim 1, characterized in that, When acquiring multi-source satellite imagery and measured water quality data of the target water body, the acquired imagery and measured water quality data are preprocessed and matched for spatiotemporal consistency to construct a dataset for model training and validation.

3. The hyperspectral / multispectral synergistic chlorophyll a inversion method based on feature-level residual learning as described in claim 1, characterized in that, The process of training a regularized regression model using multispectral broadband data to obtain the residual error vector is as follows: Broadband remote sensing reflectance is extracted from the multispectral image data in the training set, and the normalized differential chlorophyll index is calculated; the broadband reflectance and the calculated normalized differential chlorophyll index are concatenated to form a multispectral feature matrix; the multispectral feature matrix is ​​used as the input variable, and the corresponding measured chlorophyll a concentration data is used as the target variable to train the regularized regression model. After training, the multispectral feature matrix is ​​input into the trained regularized regression model to output the baseline predicted value of chlorophyll a concentration. The difference between the measured concentration and the baseline predicted value is calculated to obtain the residual error vector.

4. The hyperspectral / multispectral synergistic chlorophyll a inversion method based on feature-level residual learning as described in claim 1, characterized in that, The process of training a multivariate dimensionality reduction regression model using hyperspectral high-frequency derivative features is as follows: A continuous narrow-band matrix of the red-edge region is extracted from the hyperspectral image data in the training set; the first-order and second-order geometric gradient features of this matrix are calculated along the spectral dimension; the first-order and second-order derivative features are then column-wise concatenated to form a hyperspectral derivative feature matrix; the hyperspectral derivative feature matrix is ​​used as the input variable, and the residual error vector is used as the target variable to train the multivariate dimensionality reduction regression model.

5. The hyperspectral / multispectral synergistic chlorophyll a inversion method based on feature-level residual learning as described in claim 1, characterized in that, The chlorophyll a concentration prediction process for each pixel is as follows: A panoramic multispectral and hyperspectral remote sensing image of the target water body with unknown concentrations is acquired. The acquired image is preprocessed and matched for spatiotemporal consistency. All valid pure water body pixels are traversed, and the multispectral feature matrix of each pixel is input into a trained regularized regression model to obtain the baseline predicted value of chlorophyll a concentration for each pixel. The hyperspectral derivative feature matrix of each pixel is input into a trained multivariate dimensionality reduction regression model to obtain the high-frequency residual correction value for each pixel. The sum of the baseline predicted value of chlorophyll a concentration and the high-frequency residual correction value for each pixel is used as the initial collaborative prediction value. A soft threshold function constraint is applied to this value to obtain the final predicted value of chlorophyll a concentration for each pixel, and a high-precision collaborative chlorophyll a inversion map of the target water body is output.

6. The hyperspectral / multispectral synergistic chlorophyll a inversion method based on feature-level residual learning as described in claim 1, characterized in that, When acquiring multi-source satellite images and measured water quality data of the target water body, the hyperspectral image is thresholded using an improved normalized differential water index to generate an initial hyperspectral water body mask, and the multispectral image is generated using a red band reflectance threshold combined with connected component analysis.

7. The hyperspectral / multispectral synergistic chlorophyll a inversion method based on feature-level residual learning as described in claim 1, characterized in that, When acquiring multi-source satellite images and measured water quality data of the target water body, atmospheric correction and radiometric calibration are performed on the acquired multi-source satellite images to obtain remote sensing reflectance data.

8. The hyperspectral / multispectral synergistic chlorophyll a inversion method based on feature-level residual learning as described in claim 1, characterized in that, When acquiring multi-source satellite imagery and measured water quality data of the target water body, a continuous filtering fidelity strategy is adopted, and spatial median filtering and spectral smoothing filtering are continuously applied to the extracted hyperspectral remote sensing reflectance data cube.

9. The hyperspectral / multispectral synergistic chlorophyll a inversion method based on feature-level residual learning as described in claim 1, characterized in that, The process of performing soft threshold function constraint processing is as follows: Set the linear threshold and physical limit extreme value of the target water body, take the sum of the baseline predicted value of chlorophyll a concentration of each pixel and the high frequency residual correction value as the initial co-predicted value, remove the negative values ​​in the initial co-predicted value, and then perform soft threshold function constraint processing on it to smoothly compress the initial predicted value greater than the linear threshold and map it to the range of physical limit extreme value, so as to obtain the final predicted value of chlorophyll a concentration of each pixel, and output a high-precision co-chlorophyll a inversion map of the target water body panorama.

Images