Lake dissolved organic carbon remote sensing method integrating environmental characteristics

Abstract

The present application relates to a kind of lake dissolved organic carbon remote sensing methods of comprehensive environmental characteristics, comprising: obtaining satellite remote sensing data and calculating its equivalent reflectivity using measured water reflectivity;Extracting drainage basin boundary, obtaining attribute parameter data in drainage basin, including meteorological data, lake water parameter, topographic data and human data;Measured sample is classified using different classification methods, respectively for the correlation analysis of lake DOC concentration and all possible band ratio of satellite remote sensing data under different classification situations, determine the classification mode with the highest DOC and band ratio correlation, adopt the classification mode and classify lake into A and B type;For A type lake and B type lake, respectively using different parameters as input, DOC concentration as output constructs and trains the multilayer backtracking neural network model of corresponding type lake, selects the model with the optimal DOC simulation accuracy as the DOC estimation model of corresponding type.The method of the present application can be applied to the synchronous acquisition of DOC concentration of different types of lakes in large regional scale.

Description

Technical Field

[0001] The present invention relates to the fields of lake water optical remote sensing, water quality dynamic assessment and carbon cycle monitoring, and in particular to a lake dissolved organic carbon remote sensing method based on comprehensive environmental characteristics. Background Art

[0002] Dissolved organic carbon (DOC) in lakes can affect phytoplankton growth by absorbing ultraviolet-visible light and degrade water quality through aerobic decomposition. As the largest organic carbon reservoir, it also directly influences greenhouse gas emissions from lakes. Simultaneous monitoring of lake DOC over large spatial scales using remote sensing technology is crucial for water quality control and carbon cycle assessment. Remote sensing of lake DOC represents a regional extension of marine water research. Reported marine DOC remote sensing algorithms can generally be divided into two categories: the first utilizes the blue-green reflectance ratio for clear waters and the red-blue reflectance ratio for turbid waters; the second utilizes remote sensing to retrieve the light-sensitive colored dissolved organic matter (CDOM) in DOC and then estimates DOC from CDOM using empirical formulas. However, the complex composition of lake DOC, its interconversion with other carbon components, and its dynamic changes influenced by environmental conditions pose significant challenges to simultaneous remote sensing of DOC over large spatial scales. Although lake DOC can be traced back to the 1980s, reported DOC remote sensing algorithms are significantly regional, and most typically use the red-green band reflectance ratio or the near-infrared-red band reflectance ratio. To develop universal DOC remote sensing algorithms applicable to large regions, some studies have adopted methods such as: ① estimating the spectral slope of CDOM; ② incorporating environmental parameters such as water temperature into the model; ③ classifying water bodies based on environmental parameters and then constructing hybrid algorithms for different types; and ④ employing deep learning algorithms. With the accumulation of measured sample sizes, hybrid algorithms and deep learning have become widely used for remote sensing of environmental variables directly related to DOC concentrations, such as chlorophyll (Chl-a) and total suspended matter (TSM). However, their application to large-scale lake DOC remote sensing has yet to be reported. Summary of the Invention

[0003] The purpose of the present invention is to provide a lake dissolved organic carbon remote sensing method with universal comprehensive environmental characteristics.

[0004] In order to achieve the above technical objectives, the present invention adopts the following scheme:

[0005] A remote sensing method for lake dissolved organic carbon that integrates environmental characteristics, including:

[0006] Obtain satellite remote sensing data and calculate its equivalent reflectance using the measured water reflectance;

[0007] Extract the basin boundary and obtain the attribute parameter data within the basin, including meteorological data, lake water parameters, topographic data and humanities data;

[0008] The measured samples were classified into different classification methods, including oligotrophic / eutrophic, shallow / deep water, and freshwater / saltwater. Correlation analysis was performed between the DOC concentration of lakes under different classification conditions and all possible band ratios of satellite remote sensing data. The classification method with the highest correlation between DOC and band ratio was determined, and the lakes were classified into Class A and Class B using this classification method.

[0009] For Class A lakes and Class B lakes, the reflectance of each band + the band ratio with the highest correlation, the attribute parameter data in the basin, and the reflectance of each band + the band ratio with the highest correlation + the attribute parameter data in the basin were used as input, and DOC concentration was used as output to construct and train a multi-layer back-propagation neural network model for the corresponding type of lake. The model with the best DOC simulation accuracy was selected as the DOC estimation model for the corresponding type.

[0010] Obtain satellite remote sensing data of the lake to be measured and use the measured water reflectivity to calculate its equivalent reflectivity. After classifying the lake into Class A or Class B, select the corresponding DOC estimation model to estimate the DOC concentration.

[0011] As a preferred implementation, the satellite remote sensing data is OLCI / Sentinel-3A data.

[0012] As a preferred implementation, the satellite remote sensing data is subjected to atmospheric precision correction to calculate the equivalent reflectivity, and the atmospheric precision correction algorithm uses the MUMM method.

[0013] As a preferred embodiment, the attribute parameter data include population density, rainfall, vegetation index NDVI and slope average value in the basin; as well as lake water depth, elevation, wind speed and temperature average value within the lake surface space.

[0014] As a preferred implementation, when calculating the NDVI average and rainfall average, the annual average values ​​of different locations in the basin are first calculated based on the daily data, and then the basin average value is calculated based on the annual average value; the slope average value is calculated based on the DEM data.

[0015] As a preferred embodiment, the measured samples are divided into oligotrophic / eutrophic types according to chlorophyll a concentration, divided into shallow water / deep water types according to water depth, and divided into fresh water / salt water types according to conductivity.

[0016] As a preferred embodiment, satellite remote sensing data of the lake to be measured is obtained and its equivalent reflectivity is calculated using the measured water reflectivity, and the lake is classified as freshwater or saltwater based on the conductivity of 2000ms / cm.

[0017] As a preferred embodiment, for freshwater lakes, a multi-layer back-propagation neural network model is established with the reflectance of each band + the band ratio with the highest correlation as input and the DOC concentration as output;

[0018] For saltwater lakes, a multi-layer back-propagation neural network model was established with the attribute parameter data in the watershed as input and DOC concentration as output.

[0019] As a preferred embodiment, for freshwater lakes, a multi-layer back-propagation neural network model with three hidden layers is established, and the activation functions of the three hidden layers are purelin, logsig, and logsig respectively;

[0020] For saltwater lakes, a multi-layer back-propagation neural network model with three hidden layers is established, and the activation functions of the three hidden layers are tansig, tansig, and purelin respectively.

[0021] As a preferred embodiment, the method also includes applying the estimation model of Class A lakes to satellite remote sensing data to estimate the DOC concentration of Class A lakes in a large area, and performing monthly averaging and lake average calculations; applying the estimation model of Class B lakes to satellite remote sensing data to estimate the DOC concentration of Class B lakes in a large area, and performing monthly averaging and lake average calculations; and finally mapping the spatial distribution of DOC concentrations in lakes over a large area.

[0022] This study addresses the significant differences in DOC source composition and bio-optical properties between different lake types. By attempting to distinguish lake types using factors such as nutrient levels, water depth, and salinity, the authors ultimately discovered significant differences in DOC characteristics between freshwater and saltwater lakes. They then analyzed the correlation between DOC concentrations in freshwater and saltwater lakes and all possible band ratios, identifying the photosensitive bands for DOC in the two lake types. Finally, the authors experimented with different independent variables (CDOM, band reflectance, and environmental attributes) and algorithms (empirical methods, band ratio methods, and deep learning), ultimately selecting a deep learning model with the highest inversion accuracy for remote sensing of DOC concentrations in freshwater and saltwater lakes.

[0023] Compared with the prior art, the present application has the following advantages: ① the mixed algorithm can realize high-precision remote sensing inversion of the DOC concentration of the lake, and the average absolute percentage error is 18.16%; ② the algorithm is not sensitive to the error of atmospheric correction, and the remote sensing precision of the DOC is basically consistent when the remote sensing reflectivity error is different; ③ the salinity of the lakes in the country and the satellite reflectivity data are easy to obtain, which makes the algorithm applicable to the synchronous acquisition of the DOC concentration of different types of lakes in the country. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 . Remote sensing model selection of DOC concentration of different types of lakes

[0025] Figure 2 . Precision verification of model simulation of lake DOC concentration DETAILED DESCRIPTION

[0026] The technical solutions of the present application are further described below in combination with the description of the drawings and the specific embodiments.

[0027] The present application relates to a lake DOC remote sensing method combining the characteristics of a basin and the optics of a water body, and specifically comprises the following steps:

[0028] (1) Atmospheric correction and cutting of remote sensing data. The present application downloads the L1B level data of OLCI / Sentinel-3A from the website of the European Space Agency, and then uses the SeaDas 7.4 software to sequentially perform radiation calibration, atmospheric fine correction and image cutting. The atmospheric fine correction algorithm uses the MUMM method, which assumes that the remote sensing reflectivity ratio of the 765 nm and 865 nm bands is constant, instead of assuming that the off-water brightness of the near-infrared band is zero, and is more suitable for turbid water bodies such as lakes.

[0029] (2) Calculation of satellite band equivalent reflectivity. In order to make the DOC remote sensing algorithm constructed by the present application applicable to OLCI / Sentinel-3A satellite data, the present application first calculates the equivalent reflectivity of each visible band of OLCI / Sentinel-3A according to the spectral response function of each band of OLCI / Sentinel-3A, from the measured water body reflectivity spectrum, and the calculation formula is as formula (1):

[0030]

[0031] In the formula, f(λ) is the spectral response function of each band of OLCI / Sentinel-3A; R field (λ) is the water body reflectivity spectrum obtained by field observation; L(λ) is the solar irradiance at the average distance from the earth; λ is the integral calculation band range, taking 380-800 nm; R rs (λ) is the calculated equivalent reflectivity of each band of OLCI / Sentinel-3A.

[0032] (3) Extraction of lake basin boundaries and attributes. The HydroLAKES dataset determines the national water area larger than 20 km. 2 The lakes were then identified by combining the digital elevation model (DEM) and the HydroRIVERS dataset to determine the rivers flowing into each lake. The DEM was then subjected to operations such as depression filling, flow direction extraction, flow extraction, river network generation, river linking, and watershed extraction to generate the watershed boundaries of each lake's inflowing river. Based on the extracted watershed boundaries, average values ​​of population density, rainfall, vegetation index (NDVI), and slope were extracted, specifically:

[0033] Average rainfall: The average rainfall of the entire basin is calculated based on the rainfall at different locations within the basin (the rainfall at different locations is different). Each lake has a corresponding basin average rainfall. When calculating, the annual average of different locations within the basin is calculated based on the daily data, and then the basin average is calculated based on the annual average.

[0034] Average slope: Calculate the slope of each pixel based on the 30m resolution DEM, and then calculate the average value of the watershed;

[0035] NDVI average value: Similar to rainfall, the annual average value of different locations in the basin is calculated based on daily data, and then the basin average value is calculated based on the annual average value.

[0036] Population density: This example uses the basin average value calculated directly using the 2015 population density data.

[0037] At the same time, the average wind speed and air temperature within the lake surface space are extracted, and the lake depth and elevation are obtained from the HydroLAKES dataset.

[0038] (4) Analysis of the bio-optical characteristics of lake DOC. The measured samples were divided into oligotrophic / eutrophic, shallow water / deep water, and freshwater / saltwater types according to Chl-a concentration (10 mg / L), water depth (6.0 m), and conductivity (2000 ms / cm). The bio-optical characteristics of DOC water bodies under different classifications were compared and correlated with all possible band ratios of OLCI / Sentinel-3A. When divided into freshwater / saltwater types, the correlation between DOC and band ratio was high: for freshwater, the highest correlation coefficient was R rs (665) / R rs (490) than (r = 0.63), the lowest is R rs (490) / R rs (620) than (r = -0.56); for salt water, the highest correlation coefficient is R rs (620) / R rs (400) than (r = 0.66), the lowest is Rrs (400) / R rs (620) (r = -0.66).

[0039] (5) Selection of freshwater lake DOC remote sensing model. Figure 1 The neural network shown uses the reflectance and sensitive band ratio of each band, lake basin attributes, and both as inputs, and DOC as output. A grid search is performed to determine the optimal neural network structure for three scenarios and compare DOC simulation accuracy. Due to eutrophication and other factors, DOC concentrations in freshwater lakes often exhibit high spatiotemporal dynamics. When using the reflectance and sensitive band ratio of each band, the mean absolute percentage error (MAPD) ​​of DOC simulation is 24.10% and the root mean square error (RMSE) is 1.49 mg / L. The corresponding neural network is suitable for remote sensing inversion of DOC in freshwater lakes, namely, Equation (2):

[0040]

[0041] Where f1, f2, and f3 are the activation functions of the three hidden layers of the neural network; purelin and logsig are the corresponding function names.

[0042] (6) Selection of saltwater lake DOC remote sensing model. Figure 2 The neural network shown in the figure uses the reflectance and sensitive band ratio of each band, lake basin attributes, and both as inputs, and DOC as output. A grid search is performed to determine the optimal neural network structure for three scenarios and compare DOC simulation accuracy. Only when lake basin attributes are used as input does DOC simulation accuracy reach high levels, with a MAPD of 9.28% and a root mean square error (RMSE) of 4.09 mg / L. This corresponding neural network is suitable for remote sensing DOC inversion of saltwater lakes, as shown in Equation (2):

[0043]

[0044] Where f1, f2, and f3 are the activation functions of the three hidden layers of the neural network; tangsig and purelin are the corresponding function names.

[0045] (8) Remote sensing of DOC concentration in lakes across China. The trained formula (2) was applied to the OLCI / Sentinel-3A water reflectance to estimate the DOC concentration in freshwater lakes across China, and monthly and lake-averaged calculations were performed. The trained formula (3) was applied to the OLCI / Sentinel-3A water reflectance to estimate the DOC concentration in saltwater lakes across China, and monthly and lake-averaged calculations were performed. Finally, the spatial distribution of DOC concentration in lakes across China was mapped.

Claims

1. A remote sensing method for dissolved organic carbon in lakes based on comprehensive environmental characteristics, characterized by: include: Obtain satellite remote sensing data and calculate its equivalent reflectance using the measured water reflectance; Extract the basin boundary and obtain the attribute parameter data within the basin, including meteorological data, lake water parameters, topographic data and humanities data; The measured samples were classified into different classification methods, including oligotrophic / eutrophic, shallow / deep water, and freshwater / saltwater. Correlation analysis was performed between the DOC concentration of lakes under different classification conditions and all possible band ratios of satellite remote sensing data. The classification method with the highest correlation between DOC and band ratio was determined, and the lakes were classified into Class A and Class B using this classification method. For Class A lakes and Class B lakes, the reflectance of each band + the band ratio with the highest correlation, the attribute parameter data in the basin, and the reflectance of each band + the band ratio with the highest correlation + the attribute parameter data in the basin were used as input, and DOC concentration was used as output to construct and train a multi-layer back-propagation neural network model for the corresponding type of lake. The model with the best DOC simulation accuracy was selected as the DOC estimation model for the corresponding type. Obtain satellite remote sensing data of the lake to be measured and use the measured water reflectivity to calculate its equivalent reflectivity. After classifying the lake into Class A or Class B, select the corresponding DOC estimation model to estimate the DOC concentration.

2. The method according to claim 1, characterized in that The satellite remote sensing data is OLCI / Sentinel-3A data.

3. The method according to claim 1 or 2, characterized in that The equivalent reflectivity of the satellite remote sensing data is calculated after atmospheric precision correction, and the atmospheric precision correction algorithm uses the MUMM method.

4. The method according to claim 1, wherein The attribute parameter data include population density, rainfall, vegetation index NDVI and slope average value in the basin; as well as lake water depth, elevation, wind speed and temperature average value within the lake surface space.

5. The method according to claim 4, characterized in that When calculating the average NDVI and rainfall values, the annual average values ​​at different locations in the basin are first calculated based on the daily data, and then the basin average value is calculated based on the annual average values; the slope average value is calculated based on the DEM data.

6. The method according to claim 1, characterized in that According to chlorophyll a The measured samples were divided into oligotrophic / eutrophic types according to concentration, shallow water / deep water types according to water depth, and fresh water / salt water types according to conductivity.

7. The method according to claim 1, characterized in that Obtain satellite remote sensing data of the lake to be measured and calculate its equivalent reflectivity using the measured water reflectivity. Use the conductivity of 2000ms / cm as the boundary to classify the lake as freshwater or saltwater.

8. The method according to claim 7, characterized in that For freshwater lakes, a multi-layer back-propagation neural network model was established with the reflectance of each band plus the ratio of the band with the highest correlation as input and DOC concentration as output; For saltwater lakes, a multi-layer back-propagation neural network model was established with the attribute parameter data in the watershed as input and DOC concentration as output.

9. The method according to claim 7, characterized in that For freshwater lakes, a multi-layer back-propagation neural network model with three hidden layers was established, and the activation functions of the three hidden layers were purelin, logsig, and logsig, respectively. For saltwater lakes, a multi-layer back-propagation neural network model with three hidden layers is established, and the activation functions of the three hidden layers are tansig, tansig, and purelin respectively.

10. The method according to claim 1, characterized in that It also includes applying the estimation model for Class A lakes to satellite remote sensing data to estimate the DOC concentration of Class A lakes over a large area, and performing monthly averaging and lake average calculations; applying the estimation model for Class B lakes to satellite remote sensing data to estimate the DOC concentration of Class B lakes over a large area, and performing monthly averaging and lake average calculations; and finally mapping the spatial distribution of DOC concentration in lakes over a large area.

Images