Mangrove plant functional character parameter inversion method based on time sequence Sentinel-2 image phenology and spectral characteristics

Abstract

The invention discloses a mangrove plant functional character parameter inversion method based on time sequence Sentinel-2 image phenology and spectral characteristics, and relates to the field of obtaining mangrove plant functional character parameters through a remote sensing technology. Screening Sentinel-2L2A images to form a time sequence image data set; calculating three types of feature variables for each image pixel by pixel, and constructing an input feature set; constructing a sample data set based on the input feature set and the field actually-measured functional character parameters; performing hyper-parameter optimization on the GPR model by a Bayesian optimization algorithm; performing GPR model training by using the optimized hyper-parameters; the contribution degree of each characteristic variable is quantitatively evaluated through the SHAP; and drawing the mangrove plant functional character parameters in the research area by using the first several characteristics with the contribution degree sum greater than 90% and the trained mangrove plant functional character parameter estimation model. According to the method, phenological characteristics are considered, and mangrove plant functional character parameters can be accurately inverted.

Description

Technical Field

[0001] The present invention relates to the field of application of remote sensing technology, and in particular to the field of obtaining functional trait parameters of mangrove plants through remote sensing technology. Background Art

[0002] As an important part of the marine ecosystem, mangroves play an important role in maintaining the ecological balance of the coastal zone and preserving marine biodiversity. Mangrove plants are the most important component of the mangrove ecosystem, and their functional trait parameters can objectively characterize the response characteristics and adaptability of mangrove plant communities to changes in the external environment. Against the backdrop of increasing anthropogenic activities and global changes, accurate estimation of mangrove plant functional trait parameters will help understand and predict how the structure of mangrove plant communities and even the ecological functions of mangroves (such as productivity and stability) respond to increasing disturbance factors (such as sea level rise and frequent extreme weather conditions). It will also help formulate policies for the management and protection of mangrove ecosystems and help achieve global sustainable development, thereby playing an important role in biodiversity protection, ecological balance and ecosystem services of the entire marine ecosystem.

[0003] In recent years, remote sensing technology has become an effective method to accurately obtain the functional trait parameters of mangrove plants. At present, the remote sensing acquisition method of mangrove plant functional trait parameters is to use the basic biophysical principles of radiation transfer process and plant spectroscopy, which emphasizes the strong correlation between functional trait parameters and canopy reflectance spectra. However, the function and morphology of mangrove plants are limited by resource availability, and there is a predictable relationship between their phenology and functional trait parameters. The existing remote sensing technology for estimating mangrove plant functional trait parameters ignores the intrinsic connection between mangrove plant phenology and functional trait parameters, and no specific means are given. Summary of the invention

[0004] In view of the defects of the prior art, the present invention aims to fully consider the intrinsic relationship between mangrove plant phenology and functional trait parameters, and to provide a mangrove plant functional trait parameter inversion method based on the phenology and spectral characteristics of time-series Sentinel-2 images and the Gaussian process regression algorithm. This method can be used in mangrove growth areas to obtain accurate and high-resolution distribution of mangrove plant functional traits.

[0005] To achieve the above objectives, the technical solution of the method for inverting mangrove functional trait parameters based on the phenological and spectral characteristics of time-series Sentinel-2 images is as follows:

[0006] The method for inverting mangrove functional trait parameters based on the phenological and spectral characteristics of time-series Sentinel-2 images includes the following steps:

[0007] Step 1: Use the Google Earth Engine (GEE) cloud platform to select Sentinel-2L2A data with less cloud cover in the study area from 2022 to 2023 to construct a time series image dataset.

[0008] Step 2: Based on the GEE cloud platform and time series image datasets, three types of feature datasets are calculated for each image pixel by pixel: band feature set, vegetation index feature set, and texture feature set to construct the input feature set.

[0009] Step 3: Based on the input feature set and the field measured functional trait parameters, generate "feature-measured functional trait parameter" data pairs and construct a sample data set.

[0010] Step 4: Use the Bayesian optimization algorithm to optimize the hyperparameters of the Gaussian Process Regression (GPR) model, and use the root mean squared error (RMSE) to evaluate the results and determine the optimal hyperparameters.

[0011] Step 5: Use the optimized hyperparameters to train the GPR model and use leave-one-out cross validation to evaluate the results. The model evaluation indicators are the coefficient of determination (R 2 ), RMSE, mean absolute error (MAE) and normalized root mean square error (NormalizedRMSE, NRMSE).

[0012] Step 6: Use Shapley Additive Explanations (SHAP) to quantitatively evaluate the contribution of each feature variable.

[0013] Step 7: Use the first few features with a total contribution greater than 90% and the trained mangrove functional trait parameter estimation model to map the mangrove functional trait parameters in the study area.

[0014] The beneficial effect of the present invention is that compared with the prior art, the present invention provides an estimation method based on the phenology and spectral characteristics of time-series Sentinel-2 images, which realizes the accurate and rapid inversion and mapping of the functional trait parameters of mangrove plants. The present invention first uses the GEE cloud platform to screen Sentinel-2 images with good quality, and calculates the input feature set image by image and pixel by pixel. Then, a sample data set is constructed in combination with the functional trait parameters measured in the field. Next, the GPR model is hyperparameter optimized, the inversion model is trained using the hyperparameter optimized GPR model, and the result is evaluated using leave-one-out cross validation. Finally, SHAP is used to select the top few features with large contributions and the estimation model to map the functional trait parameters of mangrove plants. The present invention fully considers the intrinsic connection between the phenology of mangrove plants and the functional trait parameters, and uses it as an important training feature of the inversion model, thereby improving the accuracy of the inversion model, which is of great significance to the accurate inversion and mapping of the functional trait parameters of mangrove plants. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present invention, the following is a brief introduction to the implementation details with drawings. The following drawings constitute part of the specification and are used together with the implementation cases of the present invention to explain the present invention, but do not constitute a limitation of the present invention. In the drawings:

[0016] Figure 1 This is a screenshot of the image screening results;

[0017] Figure 2 It is a schematic diagram of the GPR model framework;

[0018] Figure 3 This is the spatial distribution map of functional trait parameters of four mangrove species in Zhangjiangkou Mangrove National Nature Reserve. DETAILED DESCRIPTION

[0019] Step 1: Based on the GEE cloud platform, we selected images with cloud coverage less than 10 in the study area in 2022 through the mangrove vector boundary of the Zhangjiangkou Mangrove National Nature Reserve. The image dataset is Harmonized Sentinel-2MSI L2A, and the cloud coverage screening field is "CLOUDY_PIXEL_PERCENTAGE". After calculation, there are 27 images that meet the conditions in this area. The image screening results are as follows: Figure 1 shown.

[0020] Step 2: Collect band feature sets. For each image in the image set, remove the band with a spatial resolution of 60m, retain the bands with spatial resolutions of 20m and 10m, and resample the band with a spatial resolution of 20m to 10m; calculate the vegetation index feature set. For each image in the image set, calculate kNDVI and MCARI pixel by pixel, and treat them as a band and merge them into the original image; calculate the texture feature set. First, calculate the gray level co-occurrence matrix of the central pixel of each image in the image set with a sliding window of 9*9 pixels, and then calculate the entropy and autocorrelation of kNDVI and MCARI of each image based on the gray level co-occurrence matrix, and treat them as a band and merge them into the original image. Each pixel feature set of each image contains 10 bands, 2 vegetation indices and 4 texture features. The calculation formulas for kNDVI and MCARI are as follows:

[0021] kNDVI=tanh(((P nir –P red ) / 2σ) 2 );

[0022] MCARI=((P red-edge -P red )-0.2*(P red-edge -P green ))*(P red-edge / P red )

[0023] Where P nir , P red , P red-edge and P green represents the reflectance of the near-infrared band, red band, red edge 1 band and green band respectively; σ controls the distance between the near-infrared band and the infrared band. Considering that mangroves are widely distributed in the southeastern coast of my country, the value is taken as the average of the reflectance of the near-infrared band and the infrared band of each image in the study area, that is, 0.5*(P nir +P red ).

[0024] Step 3: The coordinate system of the center position of the field-measured sample plot is converted to the coordinate system of the image. The three types of characteristic values ​​of each single pixel of each image at the coordinate are extracted through the coordinates of the center position of the sample plot, and the "characteristic-measured functional trait parameter" sample data set is constructed with the measured functional trait parameters of the corresponding sample plot. Among them, the measured functional trait parameters include leaf area index (LAI), chlorophyll content (Cab), dry matter content (Cm) and equivalent water thickness (Cw); the sample plot size is 10m*10m, which is consistent with the pixel size of Sentinel-2 image.

[0025] Step 4: Optimize the two types of hyperparameters in the GPR model. The optimization algorithm used is the Bayesian optimization algorithm. The two types of hyperparameters are the function penalty parameter ɑ GPR and kernel function, where ɑ GPR The value range is 0.001-10, with a step size of 0.001. There are three types of kernel functions: RBF kernel function, Matern kernel function, and RQ kernel function. The smoothing parameter of the Matern kernel function The value range is 0.1-2.5, with a step size of 0.1; the function penalty parameter ɑ of the RQ kernel function RQ The value range is 0.01-10, with a step size of 0.01. RMSE is used to evaluate the model results and determine the optimal hyperparameters.

[0026] Step 5: Use the optimized hyperparameters to train the GPR model. The basic principle of the GPR model is to obtain the posterior distribution of the function by defining the Bayesian relationship between the prior distribution and the observed data, thereby achieving modeling and prediction of the function. The framework diagram is shown in the figure. Figure 2 The algorithm implementation process is as follows:

[0027] Input: training dataset

[0028] T={(x1,y1),(x2,y2),...,(x N ,y N )}

[0029] Output: Hyperparameters of the Gaussian process regression function

[0030] a) Determine the data point as a sampling point of a Gaussian process, that is, a function of the mean function and the covariance function:

[0031] f(x)~N(m(x),k(x,x′))

[0032] Among them, m(x) is the mean function and k(x,x′) is the covariance function.

[0033] For simplicity, let the mean function be zero and the covariance function be a Gaussian kernel function:

[0034]

[0035] in and 2l 2 are all hyperparameters.

[0036] b) Determine the expression of the prediction point based on the posterior probability, that is, determine the expression to be solved. The distribution expression of the predicted value is:

[0037]

[0038] To make predictions, the Gaussian process is obtained by calculating the conditional distribution:

[0039] f(x)|y,{x1,x2,...,x n}; x *

[0040] The functions of the mean function and covariance function of the Gaussian distribution are:

[0041]

[0042] c) Estimation of hyperparameters of Gaussian regression model based on maximum likelihood method:

[0043]

[0044] The functional trait parameters of mangrove plants measured in the field were used as the dependent variables of the model, and all three types of characteristic parameters of each image throughout the year (i.e., a total of 432 characteristic values ​​for each pixel) were used as independent variables to construct a remote sensing inversion model of the functional trait parameters of mangrove plants.

[0045] The leave-one-out cross validation was used to evaluate the results, and the model evaluation indicators were R 2 , RMSE, MAE and NRMSE, the formulas are as follows:

[0046]

[0047] Step 6: Use the SHAP method to quantitatively evaluate the contribution of each feature variable.

[0048] Step 7: Use the first few features with a contribution greater than 90% and the trained mangrove functional trait parameter estimation model to map the mangrove functional trait parameters in the study area, such as Figure 3 shown.

[0049] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for inverting mangrove plant functional trait parameters based on the phenological and spectral characteristics of time-series Sentinel-2 images, characterized in that: The following steps are involved: S1: Obtain the Sentinel-2 time series image dataset for the whole year of 2022 based on GEE; S2: Based on the GEE cloud platform and time series image dataset, 10 image bands, 2 vegetation indices, and 4 texture features based on gray-level co-occurrence matrix are calculated pixel by pixel for each image, and an input feature set is constructed. S3: Based on the input feature set and the field measured functional trait parameters, a "feature-measured functional trait parameter" sample data set is constructed. S4: Use the Bayesian optimization algorithm to optimize the hyperparameters of the GPR model and determine the optimal hyperparameters. S5: Use the optimized hyperparameters to train the mangrove functional trait parameter model and determine the remote sensing inversion model. S6: Use SHAP to quantitatively evaluate the contribution of each feature variable. S7: Use the first few features with a total contribution greater than 90% and the trained mangrove functional trait parameter estimation model to map the mangrove functional trait parameters in the study area.

2. According to claim 1, a method for inverting mangrove plant functional trait parameters based on time-series Sentinel-2 image phenology and spectral characteristics is characterized in that: In step S1, the image data source of the Sentinel-2 time-series image dataset for the whole year of 2022 is Harmonized Sentinel-2MSI L2A, and the cloud coverage rate is less than 10%.

3. According to claim 1, a method for inverting mangrove plant functional trait parameters based on time-series Sentinel-2 image phenology and spectral characteristics is characterized in that: In the steps S2 and S3, the characteristic value of each pixel of each image of the feature data set includes 10 image bands, 2 vegetation indices, and 4 texture features based on the gray-level co-occurrence matrix. Combined with the 27 remote sensing images throughout the year, each pixel contains a total of 432 time series characteristic values; the sample data set is constructed based on the characteristic data set corresponding to the four measured functional trait parameters and the geographical location in the field, and the pixels at the corresponding sample locations constitute the "four measured functional trait parameters-432 time series characteristic values" data pairs.

4. According to claim 1, a method for inverting mangrove plant functional trait parameters based on time-series Sentinel-2 image phenology and spectral characteristics is characterized in that: In steps S4 and S5, the Bayesian optimization algorithm performs hyperparameter optimization on the function penalty parameters and kernel functions in the GPR model, wherein the kernel functions include RBF kernel functions, Matern kernel functions, and RQ kernel functions. After determining the optimal hyperparameters, the GPR inversion model is trained using the sample set to respectively determine the remote sensing inversion models of the functional trait parameters of four types of mangrove plants with the time series feature set as the dependent variable.

5. According to claim 1, a method for inverting mangrove plant functional trait parameters based on time-series Sentinel-2 image phenology and spectral characteristics is characterized in that: In steps S6 and S7, the SHAP algorithm quantitatively evaluates the contribution of each characteristic variable, and uses the first few characteristics whose sum of contributions is greater than 90% and the trained mangrove plant functional trait parameter estimation model to map the mangrove plant functional trait parameters in the study area.

Images