Forest accumulated snow albedo inversion method

Abstract

The invention discloses a forest accumulated snow albedo inversion method, and relates to the technical field of remote sensing, and the method comprises the steps: selecting a forest accumulated snow bidirectional reflection model SFBR2 as a forward model; extracting a forest accumulated snow area through the ground cover type product and the accumulated snow cover product; randomly generating a plurality of sample points in the forest accumulated snow area, and extracting input features of all the sample points; based on the input feature sample and the forest snow bi-directional reflection model SFBR2, an optimization algorithm is adopted for inversion to obtain a target variable; constructing a training data set in combination with the input features and the target variables; training a random forest model based on the constructed data set, and adopting a grid search algorithm for optimization; and a long-term time sequence image set containing input features is constructed on a GEE platform, the optimized random forest model is applied to the image set, and the albedo of the black space and the white space of the forest accumulated snow is accurately inverted. According to the method, the precision and the calculation speed of an inversion result can be remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing, and particularly to a method for retrieving the albedo of forest snow cover. Background Art

[0002] Forest snow cover is a special environment widely distributed in the mid-high latitude regions of the Northern Hemisphere, formed by the snow cover in forest areas in winter. The albedo of forest snow cover is closely related to the radiation budget and carbon cycle of the earth-atmosphere system, and has an important impact on ecological processes such as photosynthesis. Accurately obtaining the albedo of forest snow cover is of great significance for improving climate change simulation and deepening the feedback mechanism between forests and snow cover. However, the current remote sensing-based albedo retrieval methods have not been fully established for the forest snow cover scenario, which significantly affects the uncertainty of forest snow cover albedo assessment.

[0003] The methods of current mainstream remote sensing satellite products often ignore the complex characteristics of forest snow cover, directly use the forest bidirectional reflection method or calculate the albedo by linearly weighting snow cover and forest endmembers. These methods fail to fully consider the radiative transfer interaction between forests and snow cover, resulting in low accuracy of the inversion results. In addition, the lookup table method and the optimization inversion method have low computational efficiency and insufficient processing ability for large-scale and long-term sequence data, and it is difficult to meet the current actual needs for efficient and accurate snow cover albedo inversion.

[0004] Currently, a variety of forest snow cover radiative transfer models have been developed. The coupled model based on GeoSail can only simulate the vertical observation reflectance and cannot simulate the multi-angle reflectance; the coupled model based on GORT can effectively simulate the multi-angle reflectance, but assumes that the forest snow cover scenario only includes three components: snow cover, canopy, and broad-leaved leaves, which has a certain difference from the actual situation. In contrast, the forest snow cover bidirectional reflection model (SFBR2) further introduces features such as terrain, soil, coniferous leaves, and intercepted snow, making the description of the forest snow cover scenario more comprehensive and the simulation effect of the multi-angle reflectance more superior. It is currently the optimal forward model for forest snow cover inversion. The SFBR2 model accurately simulates the terrain effect through coordinate system transformation, weights the anisotropic reflection characteristics of surface snow cover and soil by the snow cover area ratio, uses the canopy radiative transfer model to simulate the complex interaction process between the leaves and intercepted snow inside the canopy, and adopts the four-stream approximation algorithm to simulate the multiple scattering process between the surface and the canopy, so as to be able to simulate the multi-angle reflectance and albedo of forest snow cover simultaneously.

[0005] The forest snow albedo inversion directly adopts the method established based on the forest scene or inverses the albedo by weighting the forest and snow components, including the following types: The kernel-driven method characterizes the anisotropic scattering characteristics of the soil-vegetation system through the isotropic kernel, geometric optical kernel, and volume scattering kernel, fits the kernel coefficients by using the reflectance of multi-day clear-sky observations, and then calculates the black-sky and white-sky albedos through hemispherical integration and narrow-band conversion; The direct evaluation method establishes the training samples between the snow-free surface reflectance and the albedo through the kernel-driven method, and establishes the training samples of the snow reflectance and the albedo by using the snow progressive radiative transfer method, and then realizes the albedo inversion based on the angular grid look-up table. The optimized inversion method can invert multiple surface parameters including the albedo at the same time. It describes the anisotropic characteristics of the mixed pixel by weighting the vegetation model and the snow model, and then obtains the optimal inversion result of the albedo by adjusting the model parameters. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for inverting the forest snow albedo to solve the above-mentioned defects in the prior art.

[0007] This method is a method for inverting the forest snow albedo, aiming to realize the efficient and accurate inversion of the large-scale forest snow albedo. This method is of great significance for accurately calculating the radiation characteristics of forest snow in remote sensing, can provide key technical support for the remote sensing monitoring and parameter inversion of forest snow, and has wide application value. Specifically, it includes the following steps:

[0008] Step 1: Through literature research and model comparison, select the forest snow bidirectional reflectance model SFBR2 as the forward model to quantitatively determine the relationship between the forest snow reflectance and the albedo;

[0009] Step 2: According to the ESA's CCI land cover type product in 2020, extract evergreen coniferous forests, deciduous coniferous forests, evergreen broad-leaved forests, deciduous broad-leaved forests, and mixed forests as the main areas of forest distribution; at the same time, according to the ERA-5 snow water equivalent product, extract the areas in the Northern Hemisphere where the snow cover days exceed 1 month as the main areas of snow distribution; extract the common part of the above two areas as the main area of forest snow distribution;

[0010] Step 3: Randomly generate 2000 sample points within the forest snow-covered area, extract the reflectance of bands 1-7, solar zenith angle, solar azimuth angle, viewing zenith angle, and viewing azimuth angle corresponding to the MODIS sensor in 2005, 2010, 2015, and 2020, and screen the high-quality forest snow observations with clear skies for three consecutive days through the quality control layer, cloud mask product, and MOD10A1 snow cover product. To reduce data redundancy, only one sample data is retained for each location per month; extract the slope, aspect, day of the year, longitude, and latitude information corresponding to all reflectance samples to form 16 input feature samples including the reflectance of MODIS bands 1-7, solar zenith angle, solar azimuth angle, viewing zenith angle, viewing azimuth angle, slope, aspect, day of the year, longitude, and latitude;

[0011] Step 4: Based on the input feature samples and the forest snow bidirectional reflectance model SFBR2, use an optimization algorithm to invert the target variables: black-sky albedo and white-sky albedo;

[0012] First, establish a cost function between the SFBR2 model simulation and the satellite-observed reflectance:

[0013]

[0014] In the formula, r i obs represents the band reflectance observed by the MODIS sensor on the i-th day, and r i mod represents the band reflectance simulated by the SFBR2 model;

[0015] Then, input the default parameters into the SFBR2 model to obtain the simulated surface reflectance, and use the minimize function in the Python SciPy library to optimize the model within the valid range of parameters to minimize the cost function, thereby obtaining the optimal model parameters. Then, input the optimized parameters into the SFBR2 model to obtain the optimized inverted black-sky albedo and white-sky albedo;

[0016] Step 5: Combine the 16 input features including the reflectance of MODIS bands 1-7, solar zenith angle, solar azimuth angle, viewing zenith angle, viewing azimuth angle, slope, aspect, day of the year, longitude, and latitude, and the 2 target variables of black-sky albedo and white-sky albedo to construct a training dataset;

[0017] Step 6: Train a random forest model based on the constructed dataset, use the grid search algorithm to optimize two hyperparameters: the number of decision trees and the maximum number of features, and obtain the result when the coefficient of determination is the highest as the optimal parameter setting;

[0018] Step 7: On the GEE (Google Earth Engine) platform, integrate 16 data including MODIS band 1 - 7 reflectance, solar zenith angle, solar azimuth angle, viewing zenith angle, viewing azimuth angle, slope, aspect, day of the year, longitude, and latitude to construct a long - term time - series image collection containing 16 input features, and apply the trained and optimized random forest model to this image collection for retrieving the black - sky and white - sky albedos in the Northern Hemisphere.

[0019] Compared with the prior art, the present invention has the following advantages:

[0020] 1. The method of the present invention combines the radiative transfer model with machine learning for retrieving forest snow albedo, giving full play to the advantages of the radiative transfer model in accurate simulation and machine learning in efficient calculation, and significantly improving the accuracy and calculation speed of the retrieval results.

[0021] 2. By combining satellite - observed reflectance with the albedo retrieved by the optimization algorithm, the method of the present invention constructs a high - quality training sample data set, which significantly enhances the representativeness of the training data set and improves the accuracy of forest snow albedo retrieval compared with the method of directly simulating using the radiative transfer model. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a schematic diagram for retrieving forest snow albedo.

[0023] Figure 2 It is the distribution area of forest snow cover and sample locations in the Northern Hemisphere.

[0024] Figure 3 It is the black - sky albedo (a) and white - sky albedo (b) retrieved based on MODIS data in January 2010. DETAILED DESCRIPTION OF THE INVENTION

[0025] To make the technical means, creative features, achieved purposes, and functions of the present invention easy to understand, the present invention will be further described below in conjunction with specific embodiments.

[0026] As Figures 1 to 3 shown, this embodiment discloses a method for retrieving forest snow albedo, which combines the parameter retrieval ability of the radiative transfer model with the efficient calculation characteristics of machine learning algorithms. By constructing a high - quality training data set through the radiative transfer model, it significantly improves the generalization ability of the machine learning model in complex forest snow scenarios, thereby achieving accurate and efficient retrieval of large - scale forest snow albedo. It specifically includes the following steps:

[0027] Step 1: Through literature research and model comparison, select the forest snow bidirectional reflectance model SFBR2 as the forward model to quantitatively analyze the relationship between forest snow reflectance and albedo.

[0028] Step 2: According to the ESA CCI land cover type product in 2020, extract evergreen coniferous forests, deciduous coniferous forests, evergreen broad-leaved forests, deciduous broad-leaved forests and mixed forests as the main areas of forest distribution; at the same time, according to the ERA-5 snow water equivalent product, extract the areas in the Northern Hemisphere where the snow cover days exceed 1 month as the main areas of snow distribution; extract the common part of the above two areas as the main area of forest snow distribution;

[0029] Step 3: Randomly generate 2000 sample points within the forest snow area, extract the reflectance of bands 1-7, solar zenith angle, solar azimuth angle, observation zenith angle, and observation azimuth angle data corresponding to the MODIS sensor in 2005, 2010, 2015, and 2020, and screen the high-quality forest snow observations with clear skies for three consecutive days through the quality control layer, cloud mask product, and MOD10A1 snow cover product. To reduce data redundancy, only one sample data is retained for each location per month; extract the slope, aspect, date within the year, longitude, and latitude information corresponding to all reflectance samples to form 16 input feature samples including MODIS bands 1-7 reflectance, solar zenith angle, solar azimuth angle, observation zenith angle, observation azimuth angle, slope, aspect, date within the year, longitude, and latitude;

[0030] Step 4: Based on the input feature samples and the forest snow bidirectional reflectance model SFBR2, use an optimization algorithm to invert the target variables: black-sky albedo and white-sky albedo;

[0031] First, establish a cost function between the SFBR2 model simulation and the satellite-observed reflectance:

[0032]

[0033] where r i obs represents the band reflectance observed by the MODIS sensor on the i-th day, and r i mod represents the band reflectance simulated by the SFBR2 model;

[0034] Then, input the default parameters into the SFBR2 model to obtain the simulated surface reflectance, and use the minimize function in the Python SciPy library to optimize the model within the valid range of parameters to minimize the cost function, thereby obtaining the optimal model parameters. Then, input the optimized parameters into the SFBR2 model to obtain the optimized inverted black-sky albedo and white-sky albedo;

[0035] Step 5: Construct a training dataset by combining 16 input features including MODIS band 1 - 7 reflectance, solar zenith angle, solar azimuth angle, viewing zenith angle, viewing azimuth angle, slope, aspect, date within a year, longitude, and latitude, as well as 2 target variables of black-sky albedo and white-sky albedo;

[0036] Step 6: Train a random forest model based on the constructed dataset, and use the grid search algorithm to optimize two hyperparameters of the number of decision trees and the maximum number of features, and obtain the result when the coefficient of determination is the highest as the optimal parameter setting;

[0037] Step 7: On the GEE platform, integrate 16 data including MODIS band 1 - 7 reflectance, solar zenith angle, solar azimuth angle, viewing zenith angle, viewing azimuth angle, slope, aspect, date within a year, longitude, and latitude to construct a long-term time series image collection containing 16 input features, and apply the trained and optimized random forest model to this image collection for inverting the black-sky and white-sky albedos in the Northern Hemisphere.

[0038] The black-sky albedo and white-sky albedo inverted based on the MODIS data in January 2010 are as Figure 3 shown. Overall, the spatial distribution characteristics of the black-sky and white-sky albedos are similar, and the albedo increases with the increase of latitude, which is related to more snow cover in high-latitude regions. At the regional scale, the albedo in the East Siberia region is significantly higher than that in other regions, because the vegetation coverage in this part of the region is relatively low, and the increased reflection of more snow cover increases the forest-snow albedo.

[0039] Generally speaking, the method of the present invention can accurately invert the albedo of snow-covered forests and effectively capture the changes in forest albedo caused by snow cover.

[0040] Therefore, the above-disclosed embodiments are illustrative in all aspects and not exclusive. All changes within the scope of the present invention or within the scope equivalent to the present invention are encompassed by the present invention.

Claims

1. A method for retrieving forest snow albedo, characterized in that, It includes the following steps: Step 1: Through literature research and model comparison, select the forest snow bidirectional reflectance model SFBR2 as the forward model to quantitatively analyze the relationship between forest snow reflectance and albedo; Step 2: Extract the main areas of forest snow distribution through land cover type products and snow cover products; Step 3: Randomly generate 2000 sample points within the forest snow area, extract the reflectance of bands 1-7, solar zenith angle, solar azimuth angle, observation zenith angle, and observation azimuth angle corresponding to the MODIS sensor in 2005, 2010, 2015, and 2020, and screen high-quality forest snow observations with clear skies for three consecutive days through the quality control layer, cloud mask product, and MOD10A1 snow cover product; extract the slope, aspect, date within the year, longitude, and latitude information corresponding to all reflectance samples to form 16 input feature samples including MODIS bands 1-7 reflectance, solar zenith angle, solar azimuth angle, observation zenith angle, observation azimuth angle, slope, aspect, date within the year, longitude, and latitude; Step 4: Based on the input feature samples and the forest snow bidirectional reflectance model SFBR2, use an optimization algorithm to invert the target variables: black-sky albedo and white-sky albedo; Step 5: Combine 16 input features including MODIS bands 1-7 reflectance, solar zenith angle, solar azimuth angle, observation zenith angle, observation azimuth angle, slope, aspect, date within the year, longitude, and latitude, as well as 2 target variables, black-sky albedo and white-sky albedo, to construct a training dataset; Step 6: Based on the constructed dataset, train a random forest model, use the grid search algorithm to optimize two hyperparameters, the number of decision trees and the maximum number of features, and obtain the result when the coefficient of determination is the highest as the optimal parameter setting; Step 7: On the GEE platform, integrate 16 data including MODIS bands 1-7 reflectance, solar zenith angle, solar azimuth angle, observation zenith angle, observation azimuth angle, slope, aspect, date within the year, longitude, and latitude to construct a long-term time series image set containing 16 input features, and apply the trained and optimized random forest model to this image set for inverting the black-sky and white-sky albedos in the Northern Hemisphere.

2. The method for retrieving forest snow albedo according to claim 1, wherein In Step 2, according to the ESA CCI land cover type product in 2020, extract evergreen coniferous forest, deciduous coniferous forest, evergreen broadleaf forest, deciduous broadleaf forest, and mixed forest as the main areas of forest distribution; at the same time, according to the ERA-5 snow water equivalent product, extract the areas in the Northern Hemisphere with snow cover days exceeding 1 month as the main areas of snow distribution; extract the common part of the above two areas as the main area of forest snow distribution.

3. The method for retrieving forest snow albedo according to claim 1, characterized in that, In Step 3, to reduce data redundancy, only one sample data is retained for each location per month.

4. A method for retrieving the albedo of forest snow cover according to claim 1, characterized in that, In Step 4, first, establish a cost function between the SFBR2 model simulation and satellite observation reflectance: where r i obs represents the band reflectance observed by the MODIS sensor on the i-th day, and r i mod represents the band reflectance simulated by the SFBR2 model; Then, the default parameters are input into the SFBR2 model to obtain the simulated surface reflectance. The minimize function in the Python SciPy library is used to optimize the model within the valid range of parameters to minimize the cost function, thereby obtaining the optimal model parameters. Furthermore, the optimized parameters are input into the SFBR2 model to obtain the optimized black-sky albedo and white-sky albedo through inversion.

Images