Remote sensing method for lake group particulate organic carbon storage

Abstract

The present application relates to a kind of lake group particle organic carbon reserves remote sensing method, comprising: based on lake remote sensing reflectivity data, the wave band ratio of different wave band combinations, peak height index and normalized carbon index are calculated, and the longitude and latitude and elevation data of synchronous ground measured sampling point are extracted, the sensitive degree of each index to measured POC concentration is verified using different function forms, and sensitive index is determined;The correlation model of each sensitive index and the natural logarithm of measured POC concentration is established respectively, as the rough estimation model of POC concentration;With the POC estimated by remote sensing reflectivity and each sensitive index as input, measured POC as output, POC concentration estimation model is constructed;Obtain lake depth and lake water area data, and calculate euphotic depth based on remote sensing reflectivity data, the POC reserve in euphotic layer is estimated using the POC concentration estimation model, the POC reserve outside euphotic layer is estimated using exponential relationship, and the sum of POC reserve in and outside euphotic layer is lake POC reserve.The present application can realize the remote sensing estimation of lake group particle organic carbon reserves.

Description

Technical Field

[0001] This invention relates to the fields of limnology, water environment monitoring, and remote sensing of lake carbon cycle, specifically to a remote sensing method for particulate organic carbon storage in a lake group. Background Technology

[0002] Particulate organic carbon (POC) originates not only from algal proliferation but also from sediment resuspension; it is not only linked to the health of lake water environment and aquatic ecosystem but is also an important form of carbon storage in lakes. Therefore, it is necessary to dynamically monitor lake POC storage using satellite remote sensing data with large-area coverage and periodic revisits.

[0003] Based on ocean POC remote sensing, the reported lake POC remote sensing algorithms can be divided into three categories: (1) Type I directly estimates POC concentration from satellite band reflectance based on empirical formulas; (2) Type II first estimates chlorophyll-a (Chl-a) concentration using satellite reflectance, and then calculates POC concentration based on Chl-a; (3) Type III first extracts inherent optical quantities such as backscattering, attenuation, and absorption coefficient from satellite reflectance based on semi-analysis algorithm, and then calculates POC concentration from inherent optical quantities.

[0004] However, previous studies have only focused on remote sensing of surface POC concentration in turbid eutrophic lakes. The applicability of reported algorithms to clear oligotrophic lakes remains unknown, and approximately 67% of the global lake area is characterized by deep water (≥6 meters), low algal content, and high transparency. Compared to turbid eutrophic lakes, clear lakes typically have lower water reflectance and a Chl-a fluorescence peak around 700 nm, and usually have a higher POC:Chl-a ratio. Therefore, further development of remote sensing algorithms is needed to simultaneously monitor POC concentrations in shallow-deep, turbid-clear, and oligotrophic-eutrophic lakes. Summary of the Invention

[0005] The purpose of this invention is to provide a remote sensing method for particulate organic carbon (POC) storage in lake groups, so as to achieve three-dimensional remote sensing estimation of POC storage in lake groups over a large area.

[0006] To achieve the above technical objectives, this invention, based on national lake POC survey data and satellite synchronous remote sensing reflectance, uses linear model statistical analysis to determine sensitive indicators of the spatiotemporal variation of lake POC concentration. Then, using the POC concentration roughly estimated from the remote sensing reflectance and sensitive indicators as input, machine learning hyperparameter optimization is performed. Finally, based on the euphotic depth and empirical relationships, a three-dimensional POC reservoir in the water column is estimated by extending the two-dimensional surface POC concentration. The specific technical solution is as follows:

[0007] A remote sensing method for particulate organic carbon storage in a lake group includes:

[0008] Based on lake remote sensing reflectance data, the band ratio, peak height index and normalized carbon index of different band combinations were calculated, and the latitude, longitude and elevation data of ground synchronous measured sampling points were extracted. The sensitivity of each index to the measured POC concentration was tested using different function forms, and the sensitive indexes were determined.

[0009] Correlation models between each sensitive indicator and the natural logarithm of the measured POC concentration were established as rough estimation models for POC concentration.

[0010] Using remotely sensed reflectance and POC concentrations estimated by various sensitivity indices as inputs and measured POC concentrations as outputs, a POC concentration estimation model is constructed.

[0011] The lake depth and water area data are obtained, and the depth of the euphotic layer is calculated based on remote sensing reflectance data. The POC storage within the euphotic layer is estimated using the POC concentration estimation model, and the POC storage outside the euphotic layer is estimated using the exponential relationship. The sum of the POC storage inside and outside the euphotic layer is the lake's POC storage.

[0012] As a preferred implementation, when using remote sensing data from different sensors of the same series of satellites for remote sensing monitoring, the remote sensing reflectance of different sensor data is calculated, and then the remote sensing reflectance of one of the sensors is used as a benchmark to fit and construct a relative calibration formula. The remote sensing reflectance of other sensors is then converted using the relative calibration formula for calculation.

[0013] As a preferred embodiment, the sensitive index is determined by using the correlation coefficient to test the correlation between the index and the POC concentration. The index that is significantly correlated is the POC concentration sensitive index.

[0014] As a preferred implementation, the POC concentration estimation model is constructed using remote sensing reflectance data in the red, blue, green, and near-infrared bands.

[0015] As a preferred implementation, the POC concentration estimation model is constructed based on the random forest method.

[0016] As a preferred implementation, the depth of the true light layer is calculated using lake transparency, which is obtained based on remote sensing inversion.

[0017] In a preferred embodiment, the POC content within the true light layer is estimated by pixel-scale integration of the POC concentration estimation model.

[0018] As a preferred embodiment, estimating the POC reserves outside the true light layer using an exponential relationship includes:

[0019] Based on measured data, an exponential relationship model was constructed between the average POC concentration in the water column inside the euphotic zone and the average POC concentration in the water column outside the euphotic zone. The average POC concentration in the water column outside the euphotic zone was then used to estimate the average POC concentration in the water column outside the euphotic zone.

[0020] As a preferred implementation, the lake water area is extracted using remote sensing reflectance data combined with a normalized water index.

[0021] As a preferred implementation, the remote sensing reflectance data is acquired based on the Landsat satellite.

[0022] The present invention has the following beneficial effects:

[0023] (1) Lake POC concentration is the result of the combined effects of various factors. In order to construct a high-precision POC remote sensing model, it is necessary to select sensitive indicators of the spatiotemporal variation of POC concentration. To address this technical challenge, this invention selects sensitive indicators that affect the spatiotemporal variation of lake POC concentration by constructing a general linear model of each indicator and ln(POC) (p<0.05).

[0024] (2) To improve the robustness and convergence speed of the POC remote sensing machine learning model, the input variables need to be optimized during model training. To address this technical challenge, in addition to band reflectance, this invention also inputs the lake POC concentration, which is roughly estimated by multiple spatiotemporally variable sensitive indicators using a general linear model, along with latitude, longitude, and elevation. Considering indicators such as latitude, longitude, and elevation in addition to remote sensing reflectance can better reflect the spatiotemporal variability of lake POC concentration. Using latitude, longitude, elevation, and the roughly estimated POC concentration from sensitive indicators as input can improve the robustness and convergence speed of the machine learning model.

[0025] (3) Satellite remote sensing can only obtain the surface POC concentration of lakes, so it is necessary to construct a correlation to estimate the POC stock in the water column from the remotely sensed surface POC concentration. To address this technical challenge, this invention dynamically estimates the average depth and euphotic depth of lakes based on remote sensing data, and then estimates the POC stock in the water column inside and outside the euphotic layer, respectively. By dynamically acquiring the lake depth and euphotic layer through remote sensing, three-dimensional remote sensing estimation of the POC stock in a large-area lake group can be achieved. Attached Figure Description

[0026] Figure 1 The remote sensing accuracy of particulate organic carbon concentration in lakes across the country: (left) training set; (right) test set. Detailed Implementation

[0027] Taking lakes within my country as an example, this invention details a specific implementation method for constructing a remote sensing method for particulate organic carbon storage in lake groups, as follows:

[0028] (1) Satellite-Ground Synchronization Data Matching. Based on the latitude and longitude and sampling date of the measured sampling data, matching data for the corresponding TM / Landsat-5, ETM+ / Landsat-7, OLI / Landsat-8 and OLI / Landsat-9 satellite data are searched within a time window of ±1 day and a time window of ±3 pixels. Then, matching data with clouds (3rd bit is 1), cirrus clouds (2nd bit is 1) and thin clouds (1st bit is 1) are removed according to the pixel quality assessment field of the matching data, and the satellite-ground matching data for clear days are retained.

[0029] (2) Landsat satellite relative calibration. Different sensors in the Landsat series satellites have certain differences in terms of band swath width and spectral response function. In order to make the POC concentration retrieved by different sensors comparable, it is necessary to perform relative calibration of the band reflectance of different sensors. This invention selects to correct the band reflectance obtained by different sensors to the OLI / Landsat-8 level. Among them, the band settings of OLI / Landsat-9 are basically the same as those of OLI / Landsat-8, and no further correction is made. For TM / Landsat-5 and ETM+ / Landsat-7, the relative calibration formula is determined based on the synchronous data. For the blue, green, red and near-infrared bands, the relative calibration formula is as shown in equation (1):

[0030]

[0031] In the formula, [R cla_blue ]、[R cla_green ]、[R cla_red ] and [R cla_NIR [R] represents the reflectance of the corrected OLI / Landsat-8 in the blue, green, red, and near-infrared bands, respectively; ori_blue ]、[R ori_green ]、[R ori_red ] and [R ori_NIR [Reflectance of TM / Landsat-5 or ETM+ / Landsat-7 in the blue, green, red and near-infrared bands before correction.]

[0032] (3) Selection of POC Sensitivity Indicators for Lakes. Based on the synchronous remote sensing reflectance of different bands from the Landsat series satellites, this invention calculated the band ratio, peak height index, and normalized difference carbon index (NDCI) for different band combinations, and extracted the latitude, longitude, and elevation of all sampling points. Then, the sensitivity of each indicator to POC concentration was tested using linear, exponential, power, and logarithmic functions, respectively. This invention found that longitude, elevation, Landsat red band reflectance, and NDCI are sensitive to POC concentration.

[0033] The peak height index is calculated by subtracting the baseline value connecting the blue and red light bands from the green light band reflectance, as shown in the following formula:

[0034] H = R cla_green -0.54R cla_blue –0.46R cla_red

[0035] The specific method for testing sensitivity is to use the correlation coefficient to test the correlation between the index and POC concentration, and to see if the significance level p is less than 0.01; for factors that are significant, R0 is selected. 2 The largest fitted function.

[0036] (4) Rough estimation of lake POC concentration. POC estimated based on POC-sensitive indicators can roughly reflect the spatial distribution pattern of POC in different lakes. Therefore, this invention uses longitude, elevation, OLI / Landsat-8 blue band reflectance, and OLI / Landsat-8 red band reflectance to roughly estimate lake POC concentration. From the satellite-to-ground synchronization data obtained in step (1), the natural logarithm of different indicators and POC concentration have the relationship shown in equation (2):

[0037]

[0038] In the formula, [Lon], [DEM], [R] cla_blue ] and [R cla_red The values ​​are the longitude, elevation, OLI / Landsat-8 blue band reflectance, and OLI / Landsat-8 red band reflectance of the sampling point, respectively.

[0039] (5) POC Random Forest Hyperparameterization. The POC results estimated by the sensitive indicators are inconsistent, and the POC calculated based on longitude and elevation cannot reflect the spatial variation within the lake. Therefore, this invention uses the POC concentration roughly estimated by the sensitive indicators and the Landsat four-band reflectance (step (2)) as inputs and the measured POC concentration as output to construct a random forest model for fine remote sensing of lake POC. The number of decision trees, the maximum depth of decision trees, the minimum number of leaf nodes and the minimum number of nodes that can be divided are set to [1,20,200] (minimum, step size, maximum), [2,2,20], [2,2,10] and [1,1,15], respectively, for the optimal POC simulation model. The final determined number of decision trees, the maximum depth of decision trees, the minimum number of leaf nodes and the minimum number of nodes that can be divided are 160, 16, 2 and 1, respectively. The mean absolute percentage error, root mean square error, and relative bias of the hyperparameterized random forest on the training dataset were 12.67%, 1.11 mg / L, and -7.16%, respectively; while on the test dataset, the mean absolute percentage error, root mean square error, and relative bias were 33.48%, 2.19 mg / L, and -8.64%, respectively.

[0040] (6) Dynamic estimation of lake average depth. Based on the remote sensing reflectance obtained in step (2), the lake water area is extracted using the normalized difference water index (NDWI).

[0041] Then, the lake surface elevation for different years was obtained by combining the DEM data, and the average lake depth for different years was obtained by combining the average lake depth for the HydroLakes dataset (2015). The calculation formula is shown in equation (3):

[0042] D(x)=D(2015) + (L(x) - L(2015)) (3)

[0043] In the formula, D(x) is the calculated water depth for the corresponding year, and D(2015) is the lake water depth in the HydroLakes dataset; L(x) is the lake surface DEM monitored based on multi-source remote sensing data, and L(2015) is the lake surface DEM in 2015.

[0044] (7) Estimation of the true light layer depth of lakes. The true light layer depth of lakes is directly related to transparency, and can be estimated from the lake transparency obtained by remote sensing inversion. Based on the remote sensing reflectance in step (2), this invention uses formula (4) to calculate the water transparency, and further calculates the true light layer depth of different lakes from the average lake transparency.

[0045]

[0046] In the formula, [SDD] represents the transparency depth retrieved from remote sensing; [R] represents the transparency depth retrieved from remote sensing. cla_red [Z] represents the corrected reflectivity of OLI / Landsat-8 in the red band; eu [ ] represents the calculated eurythemal depth of the water body.

[0047] (8) Estimation of POC reserves in lakes nationwide. Satellite remote sensing can only reflect the POC concentration at the depth of the euphotic zone in lake water. Therefore, this invention estimates the POC reserves in the water body inside and outside the euphotic zone separately. Among them, the POC reserves inside the euphotic zone are estimated by pixel-scale integration based on the POC concentration retrieved from the random forest model in step (5). The POC concentration outside the euphotic zone is estimated by the POC concentration inside the euphotic zone based on the exponential relationship. The calculation formula is as shown in formula (5):

[0048]

[0049] In the formula, [S eu ] and [S out [POC] represents the POC reserves of the water column inside and outside the euphotic zone, respectively; [POC] eu ] and [POC out [Z] represents the average POC concentration of the water column inside and outside the euphotic zone, respectively; [D] represents the average lake depth (m) obtained in step (6); eu [[Area]] represents the average euphotic depth of the lake (m); [Area] represents the lake surface area obtained from NDWI in step (6); [POC] represents the average euphotic depth of the lake. out ]=a×exp(b×[POC eu ]) is the exponential relationship determined based on measured data.

Claims

1. A remote sensing method for particulate organic carbon storage in a lake group, characterized in that, include: Based on lake remote sensing reflectance data, the band ratio, peak height index and normalized carbon index of different band combinations were calculated, and the latitude, longitude and elevation data of ground synchronous measured sampling points were extracted. The sensitivity of each index to the measured POC concentration was tested using different function forms, and the sensitive indexes were determined. Correlation models between each sensitive indicator and the natural logarithm of the measured POC concentration were established as rough estimation models for POC concentration. Using remotely sensed reflectance and POC concentrations estimated from various sensitive indicators as inputs, and measured POC concentrations as outputs, a POC concentration estimation model is constructed. Data on lake depth and water area were acquired, and the depth of the euphotic layer was calculated based on remote sensing reflectance data. The POC storage within the euphotic layer was obtained by pixel-scale integral estimation of the POC concentration estimation model. An exponential relationship model between the average POC concentration in the euphotic layer and the average POC concentration in the water column outside the euphotic layer was constructed based on measured data. The average POC concentration in the water column outside the euphotic layer was estimated using the average POC concentration in the water column inside the euphotic layer. The sum of the POC storage within and outside the euphotic layer represents the lake's POC storage.

2. The method according to claim 1, characterized in that, When using remote sensing data from different sensors on the same series of satellites for remote sensing monitoring, the remote sensing reflectance of different sensor data is calculated. Then, the remote sensing reflectance of one of the sensors is used as a benchmark for fitting, and a relative calibration formula is constructed. The remote sensing reflectance of other sensors is then converted using the relative calibration formula for calculation.

3. The method according to claim 1, characterized in that, The sensitive indicators are determined by using correlation coefficients to test the correlation between the indicators and POC concentration. The indicators that are significantly correlated are the POC concentration sensitive indicators.

4. The method according to claim 1, characterized in that, The POC concentration estimation model was constructed using remote sensing reflectance data in the red, blue, green, and near-infrared bands.

5. The method according to claim 1, characterized in that, The POC concentration estimation model was constructed based on the random forest method.

6. The method according to claim 1, characterized in that, The depth of the true light layer is calculated using lake transparency, which is obtained based on remote sensing inversion.

7. The method according to claim 1, characterized in that, The lake's water area was extracted using remote sensing reflectance data combined with a normalized water index.

8. The method according to claim 1, characterized in that, The remote sensing reflectance data was obtained based on the Landsat satellite.

Images