Chlorophyll a concentration inversion method based on spatial-spectral fusion and model learning coupling

Abstract

The present invention relates to the field of remote sensing inversion of water quality parameters, and discloses a chlorophyll a concentration inversion method based on spatial-spectral fusion and model learning coupling. The method comprises the following steps: obtaining MODIS and Sentinel‑2 data of the same date, and obtaining a sample data set after preprocessing; constructing a MODIS and Sentinel‑2 spatial-spectral fusion deep learning network for chlorophyll a concentration inversion, and coupling physical constraints such as spectral response functions and image degradation models, so as to obtain fused data with MODIS spectral resolution and Sentinel‑2 spatial resolution, and providing a high-altitude spectral resolution data source for chlorophyll a concentration inversion; combining field sampling data with fused high-altitude spectral resolution data, and verifying the effectiveness of the fusion under a water quality inversion algorithm of machine learning such as a gradient boosting tree, and the results show that the processing method effectively improves the inversion accuracy of chlorophyll a concentration.

Description

Technical Field

[0001] The invention relates to the field of water quality remote sensing inversion, and in particular to a chlorophyll a concentration inversion method based on space-spectrum fusion and model learning coupling. Background Art

[0002] Chlorophyll-a (CHL-a) is a common pigment of algae in water bodies and an important indicator for assessing the degree of eutrophication of water bodies. Although the traditional CHL-a concentration monitoring method has high accuracy, it is time-consuming and labor-intensive and difficult to apply to a large area. With the development of earth observation technology, remote sensing technology has become an important means of water quality monitoring in lakes and other places. The rich satellite remote sensing data and the continuously improved inversion models have promoted the application of remote sensing monitoring of CHL-a concentration.

[0003] However, in the process of remote sensing monitoring of CHL-A concentration, the satellite data obtained by remote sensing sensors still have the problem of mutual constraints between spatial resolution and spectral resolution. For example, MODIS data has a high spectral resolution, but the low spatial resolution increases the difficulty of setting up ground verification sampling points, limiting its application in monitoring small and medium-sized lakes and rivers; Sentinel-2 data has a high spatial resolution, but the spectral information is insufficient, which affects the inversion accuracy of CHL-A concentration.

[0004] In order to solve the problem of satellite sensor spatial spectral resolution constraints and obtain remote sensing data with high spatial spectral resolution to improve inversion accuracy, many spatial spectral fusion methods under the deep learning framework have been developed in recent years. The main principle is to use the significant nonlinear representation ability of deep learning to integrate the spatial and spectral information of different remote sensing images, that is, to improve its spatial resolution while retaining the spectral information, and finally obtain a fused image that integrates the effective information of multiple images. On the one hand, the current deep learning spatial spectral fusion method is not aimed at chl-a concentration inversion, and the fused spectrum and its parameter design are not targeted; on the other hand, the existing deep learning method lacks consideration of physical models. Therefore, the present invention proposes a chlorophyll a concentration inversion method based on spatial spectral fusion and model-learning coupling. Summary of the invention

[0005] 1. Technical issues to be resolved

[0006] The purpose of the present invention is to solve the problem that the spatial resolution and spectral resolution of satellite data used for chlorophyll a concentration inversion cannot be taken into account at the same time, and proposes a chlorophyll a concentration inversion method based on spatial-spectral fusion and model learning coupling.

[0007] (II) Technical solution

[0008] The chlorophyll a concentration inversion method based on spatial-spectral fusion and model learning coupling includes the following steps:

[0009] Step 1: Obtain MODIS and Sentinel-2 data of a certain area and perform necessary preprocessing;

[0010] Step 2: Select the sensitive bands for chlorophyll a inversion based on MODIS data, and group the MODIS bands that need to be fused based on the spectral response function of MODIS and Sentinel-2 data;

[0011] Step 3: Construct MODIS and Sentinel-2 training datasets;

[0012] Step 4: Introduce the image degradation process and perform physical constraints through the loss function. On the basis of the convolutional network, add residual connections and attention mechanisms to build a spatial-spectral fusion framework that couples the physical model with deep learning.

[0013] Step 5: Obtain n valid sample data of chlorophyll a concentration at the same time as the satellite, randomly select m of them as training data, and the remaining nm samples as verification data to verify the accuracy of the subsequent inversion model;

[0014] Step 6. According to the fused reflectance data obtained in step 4 and the band setting characteristics of Sentinel-2 reflectance data, the spectral resolution, the correlation between the reflectance data and the measured chlorophyll a concentration are comprehensively considered to determine the effective reflectance bands and their band combinations required for inversion modeling. Then, the fused reflectance bands are used together with the original Sentinel-2 reflectance bands for collaborative inversion. In combination with the measured chlorophyll a concentration data and the gradient boosting tree algorithm in machine learning, an inversion model of reflectance data and measured chlorophyll a concentration is constructed.

[0015] Furthermore, in step one, the fixed area is selected as the Chaohu area; the preprocessing includes stitching, cropping, resampling, and atmospheric correction.

[0016] Furthermore, in step 2, the specific steps of selecting sensitive bands and grouping spectral response functions are as follows:

[0017] 1) According to the characteristics of MODIS data band setting, the sensitive band is selected as the band range from near infrared to visible light. Combined with the problem of missing bands in the MODIS satellite image reflectance data in Chaohu area, the MODIS reflectance bands B1 to B4 (from the first band to the fourth band, and so on) and B8 to B12 are finally selected; the Sentinel-2 reflectance bands are B1 to B8 and B8A;

[0018] 2) Using the spectral response functions of the two satellite sensors, the MODIS bands corresponding to the same band spectral response function range of Sentinel-2 are divided into the same group. If there is no corresponding Sentinel-2 band, the MODIS band is divided into the Sentinel-2 band that is spectrally closest to it. Finally, the MODIS bands are divided into five groups, namely B1 as one group, B2 as one group, B3, B10, B11 as one group, B4, B12 as one group, and B8, B9 as one group.

[0019] Furthermore, in step 3, the specific steps of constructing the training data set are:

[0020] 1) Using the correlation of MODIS bands, select the two MODIS bands with the highest correlation with the MODIS band to be fused, and these three MODIS bands are located in different groups divided in step 2, and then select the Sentinel-2 bands corresponding to these three MODIS bands, where the Sentinel-2 band corresponding to the MODIS band to be fused is used as the label, and all bands are resampled to 20m spatial resolution. The final network input is: two pairs of MODIS-Sentinel-2 high and low spatial resolution data pairs and the MODIS low spatial resolution band to be fused;

[0021] 2) To facilitate network training, all images involved were cropped into small blocks of 80 pixels × 80 pixels with a step size of 40 pixels before being input into the network.

[0022] Furthermore, the specific steps of step 4 to construct a spatial-spectral fusion model of coupled physical model and deep learning are as follows:

[0023] 1) Convolutional neural networks are used to construct three branches: one branch is used to input Sentinel-2 high spatial resolution images, one branch is used to input MODIS low spatial resolution images, and one branch is used to input the sum of the high and low score differences of the MODIS low spatial resolution image to be fused and the first two branches; each branch contains several convolution poolings, and residual connections are added to enable deeper network training. The first two branches are used to obtain high and low score difference features of different dimensions, and then the third branch is used to map the features to the MODIS low spatial resolution image to be fused. Finally, deconvolution operations are performed under the guidance of the attention mechanism.

[0024] 2) The network's loss function LOSS consists of two parts:

[0025] That is: LOSS = RMSE1 + λ*RMSE2, λ is the regularization parameter.

[0026] RMSE1 is the spatial detail loss function, and its expression is:

[0027]

[0028] Among them, y i Represents the true value of spatial information, that is, the Sentinel-2 band data value, Represents the predicted value.

[0029] RMSE2 is a spectral fidelity loss function. Considering the physical degradation process of the image, the convolutional network is used to extract the difference features of the MODIS band to be fused and the fusion result, and the fusion result is degraded to the original MODIS scale. Another loss function is added to the network for physical constraints, specifically:

[0030]

[0031] Among them, y i ′ represents the true value of the spectral information, that is, the MODIS band data value, represents the predicted value after degradation; and is connected through a degradation model consisting of a convolutional neural network.

[0032] 3) Using the training sample data obtained in step 3, the Adam optimization algorithm is used to train the network to obtain a spatial-spectral fusion framework of coupled physical model and deep learning.

[0033] Furthermore, in step 6, the specific steps of collaborative inversion of the fused reflectivity band and the Sentinel-2 reflectivity band are as follows:

[0034] 1) Use the coupled physical model in step 4 and the spatial-spectral fusion framework of deep learning to obtain the fusion results of the reflectivity of each band.

[0035] 2) Use the fused reflectivity band and the Sentinel-2 reflectivity band for collaborative inversion, and build an inversion model of the band reflectivity and the measured chlorophyll a concentration training data set obtained in step 5; Combine the input effective reflectivity data and band combination with the measured chlorophyll a concentration training data set in step 5, and obtain the collaborative inversion accuracy through the gradient boosting tree model;

[0036] Specifically, firstly, according to the characteristics of the reflectance band after fusion and the reflectance band setting of Sentinel-2, the correlation between spectral resolution, reflectance band and measured chlorophyll a concentration was comprehensively considered to determine the effective band for inversion modeling;

[0037] Secondly, the input of the gradient boosting tree model is the valid single band and band combination form, specifically: FB8 (representing the fusion result of the eighth band of MODIS, and so on, there is no such band range in Sentinel-2), FB3, FB10 (corresponding to B2 of Sentinel-2), FB4, FB12 (corresponding to B3 of Sentinel-2), FB2 (corresponding to B8A of Sentinel-2), B1 (representing B1 of Sentinel-2, and so on), B4~B8, B1 / B3 (blue-green ratio), B4 / B3 (red-green ratio), FB2 / B4 (near infrared red ratio);

[0038] Then, a chlorophyll a concentration inversion model was constructed based on the above;

[0039] 3) Use the measured chlorophyll a concentration dataset obtained in step 5 to verify the accuracy of the inversion model.

[0040] (III) Beneficial effects

[0041] The present invention discloses a chlorophyll a concentration inversion method based on spatial-spectral fusion and model learning coupling, and the improvement lies in:

[0042] 1) An error-based adaptive coefficient is defined, and the residual of the neural network solving the objective function is fed back to the neural network, which can be more stable and flexible in practical applications.

[0043] 2) The defect of some ZNN models that they cannot be stable under noise interference is corrected by the error-based adaptive coefficient negative feedback neural network, that is, the corrected ZNN model can still accurately solve the time-varying Lyapunov equation under noise interference.

[0044] 3) The parallel computing model transforms the problem into a zero-finding problem of linear equations and is used to solve the time-varying Lyapunov equations;

[0045] 4) This method effectively improves the inversion accuracy of chlorophyll a concentration and overcomes the limitations of small and medium-sized lakes and rivers that are difficult to observe using traditional water color satellites. It helps to assess their eutrophication levels and provide a reference for monitoring, management and restoration of the water ecological environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 The figure is a flow chart of an embodiment of the present invention. DETAILED DESCRIPTION

[0047] In order to enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0048] The present invention takes Chaohu Lake as an example to invert the chlorophyll a concentration of Chaohu Lake. Figure 1 As shown, the process of the embodiment of the present invention can be divided into two stages:

[0049] Phase 1: Spatial-spectral fusion

[0050] Step 1: Obtain the MODIS and Sentinel-2 reflectance data of the Chaohu area and perform preprocessing, including stitching, cropping, resampling, and atmospheric correction.

[0051] Among them, the data involved in MODIS are MOD09 reflectivity data and MODIS L1B data. The former includes 16 bands, involving three spatial resolutions of 250m, 500m and 1000m; the latter is a data product calibrated by the instrument, containing 36 bands, and pre-processed using the MCTK plug-in in ENVI; Sentinel-2 is an L1C data product, containing 13 bands, involving three spatial resolutions of 10m, 20m and 60m, and using the Sen2Cor module for atmospheric correction.

[0052] Step 2: Select the sensitive bands for chlorophyll a inversion and introduce the spectral response function to group the MODIS bands to be fused.

[0053] According to the characteristics of MODIS data band setting, the sensitive band is selected to be the band range from near infrared to visible light. Combined with the problem of missing some bands of MODIS satellite image reflectance data in Chaohu area, the MODIS reflectance bands B1~B4 (indicates from the first band to the fourth band, and so on) and B8~B12 are finally selected; the Sentinel-2 reflectance bands are B1~B8 and B8A; using the spectral response functions of the two sensors, the MODIS bands corresponding to the same band spectral response function range of Sentinel-2 are divided into the same group;

[0054] If there is no corresponding Sentinel-2 band, the MODIS band will be divided into the Sentinel-2 band that is spectrally closest to it. Finally, the MODIS band is divided into five groups, namely B1 as a group, B2 as a group, B3, B10, B11 as a group, B4, B12 as a group, and B8, B9 as a group.

[0055] Step 3: Build MODIS and Sentinel-2 training datasets.

[0056] Using the correlation of each band of MOD09 data, as shown in Table 1, select the two MOD09 data bands with the highest correlation with the MOD09 data band to be fused, and these three MOD09 data bands are all in different groups divided in step 2, and then select the Sentinel-2 bands corresponding to these three MOD09 bands, where the Sentinel-2 band corresponding to the MOD09 band to be fused is used as the label. The final network input is: two pairs of MODIS-Sentinel-2 high and low spatial resolution data pairs and the MOIDS low spatial resolution band to be fused.

[0057] The specific correlation and selection are shown in Table 1:

[0058] Table 1 Correlation of different bands of MOD09 data

[0059]

[0060]

[0061] Among them, the near-infrared B2 band has a low correlation with each band. Considering the fusion results, B17 and B18 in MODISL1B are chosen. The correlations between these two bands and B2 are 0.957 and 0.951 respectively.

[0062] The bold black lines in each row represent the two data with the highest correlation and are in different groups.

[0063] Therefore, the two pairs of high and low spatial resolution data selected for each MODIS band are shown in Table 2:

[0064] Table 2 Two pairs of high and low spatial resolution data selected for each MODIS band

[0065]

[0066] Among them, MB10 represents the tenth band of MODIS, SB2 represents the second band of Sentinel-2, and so on, and the high and low score data pairs used by the same group of bands grouped in step 1-2 are the same.

[0067] Before inputting into the network, all bands were resampled to a spatial resolution of 20 m. To facilitate network training, the images involved were cropped into small blocks of 80 pixels × 80 pixels with a step size of 40 pixels.

[0068] Step 4: Construct a spatial-spectral fusion model of coupled physical model and deep learning.

[0069] Three branches are constructed using a convolutional neural network: one branch is used to input Sentinel-2 high spatial resolution images, one branch is used to input MODIS low spatial resolution images, and one branch is used to input the sum of the high and low score differences of the MODIS low spatial resolution image to be fused and the first two branches; each branch contains several convolution poolings, and in order to enable deeper network training, residual connections are added. The first two branches are used to obtain high and low score difference features of different dimensions, and then the third branch is used to map the features to the MODIS low spatial resolution image to be fused, and finally the deconvolution operation is performed under the guidance of the attention mechanism.

[0070] The network's loss function LOSS consists of two parts.

[0071] That is: LOSS = RMSE1 + λ*RMSE2, λ is the regularization parameter.

[0072] RMSE1 is the spatial detail loss function, and its expression is:

[0073]

[0074] Among them, y i Represents the true value of spatial information, that is, the Sentinel-2 band data value, Represents the predicted value.

[0075] RMSE2 is a spectral fidelity loss function. Considering the physical degradation process of the image, the convolutional network is used to extract the difference features between the MODIS band to be fused and the fusion result, and the fusion result is degraded to the original MODIS scale. Another loss function is added to the network for physical constraints, specifically:

[0076]

[0077] Among them, y i ′ represents the true value of the spectral information, that is, the MODIS band data value, Represents the predicted value after degradation. and is connected through a degradation model consisting of a convolutional neural network.

[0078] Regarding the determination of parameter λ: the size of λ is mainly to balance the spatial and spectral information of the fusion result. The larger the λ value, the richer the spectral information of the fusion result and the worse the spatial details; conversely, the smaller the λ value, the richer the spatial details of the fusion result and the more spectral information is lost. Adjustment idea: Considering the importance of spectral information to chlorophyll a inversion, increase the λ value as much as possible until a spatial critical point is reached, that is, if the λ value is increased further at this time, the spatial information loss will increase suddenly. At this time, the λ value is the final λ value, and the bands in the same group grouped in step 1-2 share the same λ value.

[0079] Using the training sample data obtained in step three, the Adam optimization algorithm is used to train the network to obtain a spatial-spectral fusion framework of coupled physical model and deep learning.

[0080] Stage 2: Chlorophyll a concentration inversion

[0081] Step 5: Based on the chlorophyll a concentration data obtained by actual measurement around the lake, after removing the abnormal points, a total of n (120 in this embodiment) valid sample data are collected, of which 96 (80%) sample points are randomly selected as training data, and the remaining 24 (20%) sample points are used as verification data to verify the accuracy of the constructed inversion model;

[0082] Step 6. Use the fused reflectivity band and the Sentinel-2 reflectivity band for collaborative inversion to build an inversion model of the band reflectivity and the measured chlorophyll a concentration training data obtained in step 5. Combine the input effective reflectivity data and band combination with the measured chlorophyll a concentration training data set in step 5, and obtain the collaborative inversion accuracy through the gradient boosting tree model.

[0083] Firstly, according to the characteristics of the reflectance band after fusion and the reflectance band setting of Sentinel-2, the effective bands for inversion modeling were determined by comprehensively considering the spectral resolution, the correlation between the reflectance band and the chlorophyll a concentration;

[0084] Secondly, the input of the gradient boosting tree model is in the form of valid single bands and band combinations;

[0085] Specifically: FB8 (representing the fusion result of the eighth band of MODIS, and so on, there is no such band range in Sentinel-2), FB3, FB10 (corresponding to B2 of Sentinel-2), FB4, FB12 (corresponding to B3 of Sentinel-2), FB2 (corresponding to B8A of Sentinel-2), B1 (representing B1 of Sentinel-2, and so on), B4~B8, B1 / B3 (blue-green ratio), B4 / B3 (red-green ratio), FB2 / B4 (near infrared red ratio);

[0086] Then, a chlorophyll a concentration inversion model was constructed based on the above.

[0087] Finally, the accuracy of the above inversion model is verified by combining the measured chlorophyll a concentration verification data obtained in step five.

[0088] The inversion results are: R2 is 0.87, RMSE is 45.34 μg / L. It can be seen that the collaborative inversion of the fusion band and the Sentinel-2 band has a high inversion accuracy, which verifies the effectiveness of our method.

[0089] The R2 and RMSE formulas involved are as follows:

[0090]

[0091] Among them, y i represents the true value, represents the predicted value, Represents the mean of the true values.

[0092] The above shows and describes the main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments, and that various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention, and these changes and improvements fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A chlorophyll a concentration inversion method based on spatial-spectral fusion and model learning coupling is characterized by: The following steps are involved: Step 1: Obtain MODIS and Sentinel-2 data of a certain area and perform necessary preprocessing; Step 2: Select the sensitive bands for chlorophyll a inversion based on MODIS data, and group the MODIS bands that need to be fused based on the spectral response function of MODIS and Sentinel-2 data; The specific steps for selecting sensitive bands and grouping spectral response functions are as follows: 21) According to the characteristics of MODIS data band setting, the sensitive band is selected as the band range from near infrared to visible light. Combined with the problem of missing some bands of MODIS satellite image reflectance data in a certain area, the MODIS reflectance bands B1~B4, B8~B12 are finally selected; the Sentinel-2 reflectance bands are B1~B8 and B8A; Among them, B* represents the *th band; 22) Using the spectral response functions of the two satellite sensors, the MODIS bands corresponding to the same band spectral response function range of Sentinel-2 are divided into the same group; If there is no corresponding Sentinel-2 band, the MODIS band is divided into the Sentinel-2 band that is closest to it in spectrum. Finally, the MODIS band is divided into five groups, namely B1 as one group, B2 as one group, B3, B10, B11 as one group, B4, B12 as one group, and B8, B9 as one group. Step 3: Construct MODIS and Sentinel-2 training datasets; Step 4: Introduce the image degradation process and perform physical constraints through the loss function. On the basis of the convolutional network, add residual connections and attention mechanisms to build a spatial-spectral fusion framework that couples the physical model with deep learning. Step 5: Obtain n valid sample data of chlorophyll a concentration at the same time as the satellite, randomly select m of them as training data, and the remaining nm samples as verification data to verify the accuracy of the subsequent inversion model; Step 6: According to the band setting characteristics of the fused reflectance data and Sentinel-2 reflectance data obtained in step 4, the spectral resolution, the correlation between the reflectance data and the measured chlorophyll a concentration are comprehensively considered to determine the effective reflectance bands and their band combinations required for inversion modeling. Then, the fused reflectance bands are used together with the original Sentinel-2 reflectance bands for collaborative inversion. In combination with the measured chlorophyll a concentration data and the gradient boosting tree algorithm in machine learning, an inversion model of reflectance data and measured chlorophyll a concentration is constructed. The specific steps are: 61) Using the coupled physical model of step 4 and the spatial-spectral fusion framework of deep learning, the fusion results of reflectivity of each band are obtained; 62) The fused reflectivity band and the original Sentinel-2 reflectivity band were used for collaborative inversion to construct an inversion model of the band reflectivity and the measured chlorophyll a concentration training data set obtained in step 5; specifically: first, according to the setting characteristics of the fused reflectivity band and the Sentinel-2 reflectivity band, the correlation between the spectral resolution, the reflectivity band and the measured chlorophyll a concentration was comprehensively considered to determine the effective band for inversion modeling; secondly, the input of the gradient boosting tree model is the effective single band and the band combination form; specifically: FB8 represents the fusion result of the eighth band of MODIS, and so on, in Sentinel el-2 does not have this band range; FB3 and FB10 correspond to B2 of Sentinel-2; FB4 and FB12 correspond to B3 of Sentinel-2; FB2 corresponds to B8A of Sentinel-2; B1 represents B1 of Sentinel-2, and so on for B4 to B8; B1 / B3 is the blue-green ratio, B4 / B3 is the red-green ratio, and FB2 / B4 is the near-infrared red ratio; then, the chlorophyll a concentration inversion model is constructed based on the above; combined with the input effective reflectance data and band combination and the measured chlorophyll a concentration training data set in step five, the collaborative inversion accuracy is obtained through the gradient boosting tree model; 63) Use the measured chlorophyll a concentration dataset obtained in step 5 to verify the accuracy of the inversion model.

2. The chlorophyll a concentration inversion method based on spatial-spectral fusion and model learning coupling according to claim 1 is characterized in that: In step 1, the fixed area is selected as Chaohu area; the preprocessing includes stitching, cropping, resampling, and atmospheric correction.

3. The chlorophyll a concentration inversion method based on spatial-spectral fusion and model learning coupling according to claim 1 is characterized in that: In step 3, the specific steps for constructing the training data set are: 31) Using the correlation of each MODIS band, select the two MODIS bands with the highest correlation with the MODIS band to be fused, and these three MODIS are located in different groups divided in step 2, and then select the Sentinel-2 bands corresponding to these three MODIS bands, where the Sentinel-2 band corresponding to the MODIS band to be fused is used as the label, and all bands are resampled to 20m spatial resolution. The final network input is: two pairs of MODIS-Sentinel-2 high and low spatial resolution data pairs and the MOIDS low spatial resolution band to be fused; 32) To facilitate network training, all images involved were cropped into small blocks of 80 pixels × 80 pixels with a step size of 40 pixels before being input into the network.

4. The chlorophyll a concentration inversion method based on spatial-spectral fusion and model learning coupling according to claim 1 is characterized in that: In step 4, the specific steps of constructing the spatial-spectral fusion framework of coupled physical model and deep learning are as follows: 41) Use convolutional neural networks to build three branches: One branch is used to input Sentinel-2 high spatial resolution images, one branch is used to input MODIS low spatial resolution images, and one branch is used to input the sum of the high and low score differences of the MODIS low spatial resolution image to be fused and the first two branches; Each branch contains several convolution pools. In order to train the network at a deeper level, residual connections are added. The first two branches are used to obtain high and low score difference features of different dimensions, and the third branch is used to map the features to the MODIS low spatial resolution image to be fused. Finally, the deconvolution operation is performed under the guidance of the attention mechanism. 42) The network loss function LOSS consists of two parts, namely: LOSS = RMSE1 + λ * RMSE2, Among them, λ is the regularization parameter; RMSE1 is the spatial detail loss function, and its expression is: Among them, y i Represents the true value of spatial information, that is, the Sentinel-2 band data value, represents the predicted value; RMSE2 is a spectral fidelity loss function. Considering the physical degradation process of the image, the convolutional network is used to extract the difference features of the MODIS band to be fused and the fusion result, and the fusion result is degraded to the original MODIS scale. Another loss function is added to the network for physical constraints, specifically: Among them, y i ' represents the true value of the spectral information, that is, the MODIS band data value, represents the predicted value after degradation; and is connected through a degradation model consisting of a convolutional neural network; 43) Using the training sample data obtained in step 3, the Adam optimization algorithm is used to train the network to obtain a spatial-spectral fusion framework of coupled physical model and deep learning.

5. The chlorophyll a concentration inversion method based on spatial-spectral fusion and model learning coupling according to claim 1 is characterized in that: In step 5, m=80%n, nm=20%n.

Images