A remote sensing inversion method for chlorophyll a concentration targeting sparse site data

Abstract

The present invention belongs to the technical field of water quality remote sensing inversion, and specifically relates to a remote sensing inversion method for chlorophyll-a concentration targeting sparse site data, effectively solving the limitation of insufficient matching quantity between site data and remote sensing data in the remote sensing inversion of chlorophyll-a concentration; increasing the utilization rate of sparse site data and improving the inversion accuracy of chlorophyll-a concentration at the same time. The present invention obtains MSI-MODIS data pairs on the same date, and obtains a sample data set through preprocessing; constructs a spatio-temporal fusion deep residual learning network of MSI and MODIS for chlorophyll-a concentration inversion, so as to obtain time series data with the spatial resolution of the MSI sensor at the required time; on this basis, combined with the corresponding time series site chlorophyll-a concentration observation data, an inversion model is constructed through the machine learning algorithm of Extreme Gradient Boosting (XGBoost), realizing the remote sensing inversion of chlorophyll-a concentration with limited site data.

Description

Technical Field

[0001] The present invention belongs to the technical field of water quality remote sensing inversion, and specifically relates to a method for remote sensing inversion of chlorophyll-a concentration for sparse site data. Background Technique

[0002] The chlorophyll-a concentration is usually used to reveal the eutrophication state of aquatic ecosystems. Traditional water quality monitoring is restricted by time, cost, and region. Fortunately, satellite remote sensing technology can improve the monitoring ability, improve the chlorophyll-a monitoring project, and facilitate the realization of large-scale long-term dynamic monitoring. At present, using MSI data to carry out remote sensing inversion of chlorophyll-a concentration has advantages such as high accuracy and high spatial resolution, and is widely used in the monitoring research of chlorophyll-a concentration in large and medium-sized lakes and rivers.

[0003] However, previous methods often collected sufficient measured data according to the satellite overpass time for chlorophyll-a concentration inversion, and the results were not necessarily the data at the required time; when there were more collection points, the collection time span would increase, which was deviated from the satellite overpass instantaneous time and affected the inversion accuracy. Although the national control stations can collect automatically at high frequencies and are basically consistent with the satellite overpass instantaneous time, the site layout is sparse, and the data pairs composed of one-phase images and sites are insufficient, affecting the inversion accuracy.

[0004] Therefore, in order to solve the limitation of the insufficient matching quantity of site data and remote sensing data in the remote sensing inversion of chlorophyll-a concentration, considering MSI and MODIS (1d time resolution) for spatio-temporal fusion to obtain the temporal data with the spatial resolution of the MSI data sensor at the required time. In recent years, due to the rise of deep learning, many spatio-temporal fusion methods under deep learning frameworks have been developed and achieved good results in many fields. However, the fusion data obtained by the deep learning spatio-temporal fusion method has not been applied to chlorophyll-a inversion at present. Therefore, the present invention provides a new inversion method to solve this problem. Summary of the Invention

[0005] The object of the present invention is to provide a method for remote sensing inversion of chlorophyll-a concentration for sparse site data aiming at the limitation of the insufficient matching quantity of site data and remote sensing data in the remote sensing inversion of chlorophyll-a concentration.

[0006] To achieve the above technical object and reach the above technical effect, the present invention is realized through the following technical solutions:

[0007] A method for remote sensing inversion of chlorophyll-a concentration for sparse site data includes the following steps:

[0008] S1. Obtain the MSI and MODIS data of the region and perform necessary preprocessing, and the preprocessing includes mosaicking, cropping, resampling, and atmospheric correction;

[0009] S2. Select the sensitive bands for chlorophyll a inversion on the MSI data;

[0010] S3. Combine the band ranges corresponding to the MSI and MODIS data to construct the MSI-MODIS training dataset;

[0011] S4. Add residual connections on the basis of the convolutional network to construct a deep learning spatio-temporal fusion model. Conduct real experiments on the basis of obtaining the optimal parameters in the simulation experiment, so as to obtain the time series data with the spatial resolution of the MSI data sensor at the required time. This data includes the fused reflectance data and the original MSI reflectance data;

[0012] S5. Determine the bands and band combinations required for inversion modeling according to the characteristics of the MSI sensor band settings and the sensitive bands for chlorophyll a concentration inversion; Conduct subsequent joint inversion using the time series data composed of the fused reflectance data and the original MSI reflectance data;

[0013] S6. Based on the above time series data and the chlorophyll a concentration data of the regional stations in the same period, construct n + p effective sample pairs. Randomly select m + p of them as the training dataset, and construct a joint inversion model through the machine learning algorithm of the extreme gradient boosting tree; The remaining n - m sample pairs are used as the verification dataset to verify the accuracy of the constructed inversion model.

[0014] Further, in step S2, the specific steps for sensitive band selection are as follows:

[0015] According to the characteristics of the MSI data band settings, the sensitive bands are selected in the band range from visible light to near infrared. Finally, the MSI reflectance bands are selected as B1 - B8 and B8A.

[0016] Further, in step S3, the specific steps for constructing the training dataset are as follows:

[0017] 1) According to the MSI band range, select the MODIS bands corresponding to its bands. If there is no corresponding MODIS band in this MSI band range, select the MODIS band closest to its band range; The selected bands are used as the MSI-MODIS data pairs at the intermediate time T2. Then select two pairs of MSI-MODIS high and low spatial resolution data pairs at the times T1 and T3 before and after it. The MSI band corresponding to the T2 time is used as the Label. All bands are resampled to a 20m spatial resolution. The final network input is: two pairs of MSI-MODIS high and low resolution spatial data pairs at the T1 and T3 times and the MODIS band at the T2 time;

[0018] 2) For the convenience of network training, before the network input, the involved images are all cropped into small images of 80 pixels × 80 pixels, with a step size of 40 pixels.

[0019] Further, in step S4, the specific steps for constructing the spatio-temporal fusion model of deep learning and conducting simulation experiments and real experiments are as follows:

[0020] 1) Use a convolutional neural network to construct three branches. Branch one is used to input the MSI high-spatial-resolution images at T1 and T3 moments, branch two is used to input the MODIS low-spatial-resolution images at T1 and T3 moments, and branch three is used to input the sum of the difference between the MODIS low-spatial-resolution image at T2 moment and the first two branches; each branch contains several convolutional pooling operations;

[0021] 2) The loss function of the network is composed of Huber, and its expression is:

[0022]

[0023] where y i represents the true value, that is, the MSI band data value, represents the predicted value, and δ is a hyperparameter;

[0024] 3) Use the training sample data obtained in step S3, and adopt the Adam optimization algorithm to train the network to obtain a spatio-temporal fusion model based on deep learning;

[0025] 4) Use the constructed spatio-temporal fusion model based on deep learning to conduct simulation experiments to obtain the optimal parameters, obtain the reflectance image fusion result of the true MSI corresponding to the date, and conduct an evaluation; on the basis of the simulation experiment, conduct a real experiment to obtain the fusion result without the true MSI corresponding to each date, and jointly form the subsequent time-series data for joint inversion with the original MSI reflectance.

[0026] Further, in sub-step 1) of step S4, each branch contains several convolutional pooling operations. In order to train the network at a deeper level, residual connections are added and implemented through the deconvolution process.

[0027] Further, in step S5, the specific steps for determining the bands and band combinations required for inversion modeling are as follows:

[0028] 1) Use the time-series data composed of the fused reflectance data and the MSI reflectance data in sub-step 4) of step S4 to carry out subsequent joint inversion;

[0029] 2) According to the characteristics of the MSI reflectance band settings and the sensitive bands for chlorophyll a concentration inversion, determine the bands input for inversion modeling;

[0030] 3) The input of the extreme gradient boosting tree model is in the form of single bands and band combinations. Specifically, for MSI, the bands used are SB1 to SB8, where SB1 represents B1 of Sentinel-2 MSI, and so on; for the fused images, the fused bands corresponding to the above MSI band range are used, i.e., FB1 to FB8, where FB1 represents the fused band corresponding to MSI B1, and so on.

[0031] Further, in step S6, the specific steps for obtaining the training and validation data sets of the joint inversion model are as follows:

[0032] 1) Based on the time-series data provided in step S4 and the chlorophyll a concentration data of regional stations in the same period, construct n + p effective sample pairs.

[0033] 2) Among the original n sample pairs of MSI data and stations in the same period, randomly select m sample pairs as the training data set, and the remaining n - m as the basis of the test data set. Then, add p sample pairs of fused data and stations in the same period, and use the m + p sample pairs as the training data set for joint inversion. Construct a joint inversion model through the machine learning algorithm of extreme gradient boosting tree.

[0034] 3) Keep the test data set of the original n - m sample pairs unchanged to verify the accuracy of the constructed joint inversion model.

[0035] Further, in sub-step 2) of step S6, n, m, and p are all positive integers, and the ratio of m / n is controlled between 0.7 and 0.9.

[0036] The beneficial effects of the present invention are as follows:

[0037] 1. The present invention fully combines the spatio-temporal complementary information between multi-scale remote sensing data, effectively solves the limitation of insufficient matching quantity between in-situ data and remote sensing data in the remote sensing inversion of chlorophyll a concentration; increases the utilization rate of sparse in-situ data, and at the same time improves the inversion accuracy of chlorophyll a concentration.

[0038] 2. The present invention obtains MSI-MODIS data pairs on the same date, and obtains a sample data set after preprocessing; constructs a spatio-temporal fusion deep residual learning network for MSI and MODIS for chlorophyll a concentration inversion, so as to obtain time-series data with the spatial resolution of the MSI sensor at the required time; on this basis, combined with the corresponding time-series in-situ chlorophyll a concentration observation data, construct an inversion model through the machine learning algorithm of extreme gradient boosting tree (XGBoost), realizing the remote sensing inversion of chlorophyll a concentration for sparse national control station data.

[0039] 3. The present invention can play a role in waters such as large and medium-sized inland lakes or rivers with stations, which helps to evaluate the eutrophication level, provides reference for water body ecological environment monitoring, management and restoration, and has great practical value.

[0040] Of course, it is not necessary for any product implementing the present invention to achieve all of the above advantages simultaneously. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for describing the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0042] Figure 1 is the flowchart of Embodiment 1 of the present invention;

[0043] Figure 2 is the band distribution diagram of MSI and MODIS between the wavelengths of 0.4 - 0.9 μm, where MOO09 B13 - B16 (saturated in the Chaohu Lake area and shown in gray) is not included in the research scope of Embodiment 1. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0044] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some, rather than all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the protection scope of the present invention.

[0045] A remote sensing inversion method for chlorophyll a concentration targeting sparse site data includes the following steps:

[0046] S1. Obtain MSI and MODIS data of the region and perform necessary preprocessing, including mosaicking, cropping, resampling, and atmospheric correction;

[0047] S2. Select the sensitive bands for chlorophyll a inversion on the MSI data;

[0048] S3. Combine the band ranges corresponding to the MSI and MODIS data to construct an MSI - MODIS training dataset;

[0049] S4. Add residual connections on the basis of the convolutional network to construct a spatio-temporal fusion model for deep learning. Conduct real experiments on the basis of the optimal parameters obtained from the simulation experiments, so as to obtain the time-series data with the spatial resolution of the MSI data sensor at the required time. This data includes the fused reflectance data and the original MSI reflectance data;

[0050] S5. According to the characteristics of the MSI sensor band settings and the sensitive bands for chlorophyll a concentration inversion, determine the bands and band combinations required for inversion modeling; use the time-series data composed of the fused reflectance data and the original MSI reflectance data to carry out subsequent joint inversion;

[0051] S6. Based on the above time-series data and the chlorophyll a concentration data of the regional stations in the same period, construct n + p effective sample pairs, randomly select m + p of them as the training data set, and construct a joint inversion model through the machine learning algorithm of the extreme gradient boosting tree; the remaining n - m sample pairs are used as the verification data set to verify the accuracy of the constructed inversion model.

[0052] The specific embodiments of the present invention are as follows:

[0053] Embodiment 1

[0054] Taking Chaohu Lake as an example, this embodiment inverses the chlorophyll a concentration in Chaohu Lake. As Figure 1 shown, the process of this embodiment can be divided into two stages:

[0055] The first stage: Spatio-temporal fusion

[0056] Step 1-1. Obtain the MSI and MODIS reflectance data of the Chaohu Lake area and conduct preprocessing, including mosaicking, cropping, resampling, and atmospheric correction. Among them, the data involved in MODIS is the MOD09 reflectance data, including 16 bands, involving three spatial resolutions of 250m, 500m, and 1000m; Sentinel-2 MSI uses the L1C data product, which contains 13 bands, involving three spatial resolutions of 10m, 20m, and 60m, and the Sen2Cor module is used for atmospheric correction. The specific band distributions of the two sensors are as Figure 2 shown.

[0057] Step 1-2. Select the sensitive bands for chlorophyll a inversion. According to the characteristics of the MSI data band settings, the sensitive bands are selected in the band range from near-infrared to visible light. The finally selected MSI reflectance bands are B1 to B8 (indicating from the first band to the fourth band, and so on) and B8A; corresponding to the band range of MSI, and considering that the pixels covering the Chaohu Lake area are saturated in MOD09 B13 - B16 ( Figure 2 the bands shown in grayish white), the selected MODIS reflectance bands are B1 to B4 and B9. CombiningFigure 2 The band distribution of

[0058] Table 1

[0059] MSI MODIS B1 B9 B2 B3 B3 B4 B4 - B6 B1 B7 - B8, B8A B2

[0060] Step 1-3: Construct the MSI-MODIS training dataset. According to the MSI band range, select the MODIS bands corresponding to its bands (if there is no corresponding MODIS band within the MSI band range, select the MODIS band closest to its band range). This serves as the MSI-MODIS data pair at the intermediate time (T2). Then, select two pairs of MSI-MODIS high and low spatial resolution data pairs at the times before and after it (T1, T3). The MSI band corresponding to the T2 time is used as the Label. All bands are resampled to a 20m spatial resolution. The final network input is: two pairs of MSI-MODIS high and low spatial resolution data pairs at the T1 and T3 times and the MODIS band at the T2 time.

[0061] Before inputting into the network, all bands are resampled to a 20m spatial resolution. For the convenience of network training, the images involved are cropped into small images of 80 pixels × 80 pixels with a step size of 40 pixels.

[0062] Step 1-4: Construct a spatio-temporal fusion model based on deep learning. Use a convolutional neural network to construct three branches. Branch one is used to input the MSI high spatial resolution images at the T1 and T3 times. Branch two is used to input the MODIS low spatial resolution images at the T1 and T3 times. Branch three is used to input the sum of the difference between the MODIS low spatial resolution image at the T2 time and the first two branches (branch one, branch two). Each branch contains several convolutional pooling operations. At the same time, in order to train the network at a deeper level, residual connections are added and implemented through the transposed convolution process.

[0063] The loss function of the network is composed of Huber, and its expression is:

[0064]

[0065] where y i represents the true value, that is, the MSI band data value, represents the predicted value, and δ is a hyperparameter.

[0066] Using the training sample data obtained in Steps 1 - 3, the Adam optimization algorithm is used to train the network to obtain a spatio - temporal fusion model based on deep learning. Based on the optimal parameters obtained from the simulation experiment, a real experiment is carried out to obtain the time - series data with the spatial resolution of the MSI data sensor at the required time (including the fused reflectance data and the original MSI reflectance data). Table 2 shows the relevant time information and its uses of the MSI - MODIS data pairs involved in the spatio - temporal fusion process, specifically:

[0067] Table 2

[0068]

[0069]

[0070] Table 3 shows the evaluation of the fusion results of the simulation experiment, specifically:

[0071] Table 3

[0072]

[0073] Among them, RMSE, SSIM, and CC are three evaluation indicators of the fused result image, and their expressions are as follows:

[0074]

[0075]

[0076]

[0077] y i 、 represent the true value and its average value of the MSI band data respectively, represent the predicted value and its average value respectively, and C1 and C2 are constants.

[0078] The second stage: Chlorophyll a concentration inversion

[0079] Step 2-1, use the time series data composed of the fused reflectance data and MSI reflectance data in step 1-3 to carry out subsequent joint inversion. According to the characteristics of the MSI reflectance band setting and the sensitive band of chlorophyll a concentration inversion, determine the band for input inversion modeling. The input of the XGBoost model is a single band and a band combination, specifically: for MSI, the bands used are SB1~SB8 (SB1 represents B1 of Sentinel-2MSI, and so on, SB8A, SB1 / SB3 (blue-green ratio), SB4 / SB3 (red-green ratio), SB8A / SB4 (near-infrared red ratio); for the fused image, the fused band corresponding to the above MSI band range is used, that is, FB1~FB8 (FB1 represents the fused band corresponding to MSI B1, and so on), FB8A, FB1 / FB3, FB4 / FB3, FB8A / FB4.

[0080] Step 2-2: Based on the time series data mentioned in step 1-3 and the chlorophyll a concentration data of Chaohu national monitoring stations in the same period, 69 valid sample point pairs were constructed. On the basis of the original MSI data and 36 sample point pairs of the stations in the same period [29 random sample point pairs (close to 80%) were used as training data sets, and the remaining 7 (close to 20%) were used as test data sets], the fusion data and 33 sample point pairs of the stations in the same period were added, and 62 sample point pairs were used as training data sets for joint inversion. The joint inversion model was constructed through the machine learning algorithm of the extreme gradient boosting tree (XGBoost); the test data set of the original 7 sample point pairs was kept unchanged to verify the accuracy of the constructed joint inversion model. The inversion result is: R 2 The inversion accuracy of the fusion band and the MSI band is 0.88, and the RMSE is 0.87 μg / L. It can be seen that the collaborative inversion of the fusion band and the MSI band has a higher inversion accuracy, which verifies the effectiveness of our method.

[0081] The R 2 The RMSE formula is as follows:

[0082]

[0083] Among them, y i represents the true value, represents the predicted value, Represents the mean of the true values.

[0084] The preferred embodiments of the present invention disclosed above are only used to help illustrate the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific embodiments. Obviously, many modifications and variations can be made according to the content of this specification. These embodiments are selected and specifically described in this specification in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can well understand and utilize the present invention. The present invention is only limited by the claims and their full scope and equivalents.

Claims

1. A remote sensing inversion method for chlorophyll a concentration targeting sparse site data, characterized in that, It includes the following steps: S1. Obtain the MSI and MODIS data of the area and perform necessary preprocessing, including mosaicking, cropping, resampling, and atmospheric correction; S2. Select the sensitive bands for chlorophyll a inversion on the MSI data; S3. Combine the band ranges corresponding to the MSI and MODIS data to construct an MSI-MODIS training dataset; S4. Add residual connections on the basis of the convolutional network to construct a deep learning spatio-temporal fusion model, and conduct a real experiment on the basis of obtaining the optimal parameters in the simulation experiment, so as to obtain the temporal data with the spatial resolution of the MSI data sensor at the required time. This data includes the fused reflectance data and the original MSI reflectance data; The specific steps for constructing a deep learning spatio-temporal fusion model and conducting simulation experiments and real experiments are as follows: 1) Use a convolutional neural network to construct three branches. Branch 1 is used to input the MSI high-spatial-resolution images at T1 and T3 times, Branch 2 is used to input the MODIS low-spatial-resolution images at T1 and T3 times, and Branch 3 is used to input the sum of the difference between the MODIS low-spatial-resolution image at T2 time and the first two branches; Each branch contains several convolutional pooling operations; 2) The loss function of the network is composed of Huber, and its expression is: Among them, represents the true value, that is, the MSI band data value, represents the predicted value, is a hyperparameter; 3) Use the training sample data obtained in step S3 and adopt the Adam optimization algorithm to train the network to obtain a deep learning spatio-temporal fusion model; 4) Use the constructed deep learning spatio-temporal fusion model to conduct a simulation experiment to obtain the optimal parameters, obtain the fused result of the reflectance image of the real MSI on the corresponding date, and conduct an evaluation; On the basis of the simulation experiment, conduct a real experiment to obtain the fused result of the real MSI without corresponding dates, which together with the original MSI reflectance constitutes the temporal data for subsequent joint inversion; S5. Determine the bands and band combinations required for inversion modeling according to the band setting characteristics of the MSI sensor and the sensitive bands for chlorophyll a concentration inversion; Use the temporal data composed of the fused reflectance data and the original MSI reflectance data to carry out subsequent joint inversion; S6. Based on the above temporal data and the chlorophyll a concentration data of regional stations in the same period, construct n + p effective sample pairs, randomly select m + p of them as the training dataset, and construct a joint inversion model through the machine learning algorithm of the extreme gradient boosting tree; The remaining n - m sample pairs are used as the verification dataset to verify the accuracy of the constructed inversion model.

2. The remote sensing inversion method for chlorophyll a concentration targeting sparse site data according to claim 1, characterized in that, In step S2, the specific steps for sensitive band selection are as follows: According to the band setting characteristics of the MSI data, the sensitive bands are selected as the band range from visible light to near-infrared, and finally the MSI reflectance bands B1~B8 and B8A are selected.

3. The chlorophyll a concentration remote sensing inversion method for sparse site data according to claim 1, characterized in that In step S3, the specific steps for constructing the training dataset are as follows: 1) According to the MSI band range, select the MODIS band corresponding to it. If there is no corresponding MODIS band within the MSI band range, select the MODIS band closest to its band range; the selected band is used as the MSI-MODIS data pair at the intermediate time T2, and then select two pairs of MSI-MODIS high and low spatial resolution data pairs at the times T1 and T3 before and after it. The MSI band corresponding to the time T2 is used as the Label, and all bands are resampled to a 20 m spatial resolution. The final network input is: two pairs of MSI-MODIS high and low spatial resolution data pairs at the times T1 and T3 and the MODIS band at the time T2; 2) For the convenience of network training, before the network input, the images involved are cropped into small images of 80 pixels × 80 pixels, with a step size of 40 pixels.

4. The remote sensing inversion method for chlorophyll a concentration targeting sparse site data according to claim 1, characterized in that In sub-step 1) of step S4, each branch contains several convolutional pooling operations. To enable the network to be trained at a deeper level, residual connections are added and implemented through the deconvolution process.

5. The remote sensing inversion method for chlorophyll a concentration of sparse site data according to claim 4, characterized in that In step S5, the specific steps to determine the bands and band combinations required for the inversion modeling are as follows: 1) Use the time series data composed of the fused reflectance data and the MSI reflectance data in sub-step 4) of step S4 to carry out subsequent joint inversion; 2) Determine the bands input for the inversion modeling according to the characteristics of the MSI reflectance band settings and the sensitive bands for chlorophyll a concentration inversion; 3) The input of the extreme gradient boosting tree model is in the form of single bands and band combinations. Specifically, for MSI, the bands used are SB1~SB8, where SB1 represents B1 of Sentinel-2 MSI, and so on; for the fused image, the fused bands corresponding to the above MSI band range are used, that is, FB1~FB8, where FB1 represents the fused band corresponding to MSI B1, and so on.

6. The method for remote sensing inversion of chlorophyll a concentration for sparse site data according to claim 1, characterized in that, In step S6, the specific steps to obtain the training and validation data sets for the joint inversion model are as follows: 1) According to the time series data provided in step S4 and the chlorophyll a concentration data of the regional sites in the same period, construct n + p effective sample pairs; 2) Among the original n sample pairs of MSI data and the sites in the same period, randomly select m sample pairs as the training data set, and the remaining n - m as the basis for the test data set. Then, add p sample pairs of fused data and the sites in the same period, and use the m + p sample pairs as the training data set for the joint inversion. Construct the joint inversion model through the machine learning algorithm of the extreme gradient boosting tree; 3) Keep the test data set of the original n - m sample pairs unchanged to verify the accuracy of the constructed joint inversion model.

7. The method for remote sensing inversion of chlorophyll a concentration for sparse site data according to claim 6, wherein In sub-step 2) of step S6, n, m, and p are all positive integers, and the ratio of m / n is controlled between 0.7 and 0.9.

Images