Evergreen vegetation aboveground biomass remote sensing estimation method based on vegetation optical thickness

Abstract

The invention discloses a vegetation optical thickness-based evergreen vegetation aboveground biomass remote sensing estimation method. The method comprises the following steps of: acquiring remote sensing image data and preprocessing the remote sensing image data; according to the active and passive microwave downscaling model, the rough surface effective emissivity model and the Tau-Omega radiation transmission model, inverting high-spatial-resolution vegetation optical thickness distribution information, and constructing a precision evaluation scheme; and optimizing and establishing an overground biomass remote sensing estimation model to obtain the overground biomass spatial distribution of the evergreen vegetation at the optimal scale. Aiming at the problem that the biomass change on evergreen vegetation is difficult to effectively monitor through a canopy greenness vegetation index, the advantages of Sentinel-1 and SMOS satellites are utilized, and multi-source remote sensing data such as rough surface effective emissivity, surface temperature, soil texture and the like are combined; and a high-spatial-resolution vegetation optical thickness index sensitive to the change of the overground biomass of the evergreen vegetation is constructed, and accurate inversion of the overground biomass of the large-scale evergreen vegetation is realized.

Description

Technical Field

[0001] The invention relates to the field of remote sensing estimation of vegetation aboveground biomass (AGB), and in particular to a remote sensing estimation method of evergreen vegetation AGB based on vegetation optical depth (VOD). Background Art

[0002] Vegetation AGB is one of the key components of terrestrial ecosystems. It provides the necessary nutrient source and energy basis for global ecological balance, ecosystem services, biodiversity maintenance and carbon cycle processes. Using remote sensing technology to accurately estimate vegetation AGB on a large scale is one of the current research hotspots in the field of ecological environment remote sensing.

[0003] Quantification of large-scale vegetation AGB requires timely and accurate information on changes in the composition, distribution and structure of vegetation ecosystems. The restoration or destruction of vegetation cover is the most intuitive manifestation of ecosystem disturbance. For evergreen vegetation areas, changes in AGB are far more than just reflected in vegetation coverage. In the seedling / forest growth stage, the increase in vegetation canopy or canopy density is usually an important indicator of AGB changes; however, when the vegetation canopy coverage is as high as 80%-90% and the forest density is high, the change in AGB mainly comes from the contribution of wood components such as roots, stems and branches, which is mainly reflected in the change of the vertical structure of vegetation. These wood components are also the components of the vegetation ecosystem and are also involved in the material cycle, energy flow and information transmission process. They are often ignored but very important in the study of ecosystem changes. Short-wavelength optical remote sensing satellites usually obtain surface reflectance information in the early growth of young forests and dense vegetation canopies. They cannot penetrate the canopy to capture changes in wood components of evergreen vegetation. When the vegetation coverage reaches a certain level, it will no longer be sensitive to changes in its canopy coverage or greenness due to the saturation effect. In addition, optical remote sensing satellite data are also easily affected by cloudy and rainy weather, resulting in low data quality and continuity. Relatively speaking, passive microwave remote sensing satellites in low-frequency bands have strong penetration capabilities, can better monitor the radiation energy of vegetation canopies and their lower parts, and are not affected by cloud and rain conditions. They have the advantage of continuous observation of surface changes all day and all day. The VOD index derived from passive microwave satellites is a key parameter for measuring the attenuation characteristics of vegetation during microwave radiation transmission. It can not only effectively capture the changes in moisture content of vegetation photosynthetic biomass and non-photosynthetic biomass, and provide water content and structural information in addition to the physical and chemical conditions of leaves, but also has low sensitivity to the saturation effect of areas with high-density vegetation coverage. Based on the above advantages, VOD has gradually become a potential ideal indicator for describing changes in vegetation hydraulics, phenology and biomass at regional and even global scales.

[0004] However, the spatial resolution of existing passive microwave VOD products is generally coarse (~25-40km), which makes it difficult to clearly reflect the heterogeneity of vegetation growth changes at the sub-pixel scale and the differences in the response of different community types to disturbances, and thus cannot meet the needs of estimating the AGB of different evergreen vegetation at a finer scale. Active microwave data represented by Sentinel-1 satellites, with their advantages of high spatial resolution, continuous observation and free access, provide a favorable opportunity for the generation of VOD products with higher spatial resolution by combining active and passive microwave remote sensing satellites. In addition, although there are various types of methods and models that can be used for AGB estimation, it is difficult to select a general model that is suitable for vegetation growth and distribution characteristics in different regions, and there is an obvious scale effect problem in the estimation of AGB on complex heterogeneous surfaces. Therefore, it is urgent to determine the spatial scale and method model suitable for remote sensing estimation of AGB on complex heterogeneous surfaces, which is also an important basis for accurate estimation of large-scale evergreen vegetation AGB based on VOD index. Summary of the invention

[0005] Aiming at the problem that the canopy greenness vegetation index in the prior art is prone to spectral saturation effect in densely vegetation-covered areas and has low sensitivity to the AGB changes of the vertical structure of evergreen vegetation, and cannot accurately estimate the large-scale evergreen vegetation AGB, the present invention proposes a VOD-based evergreen vegetation AGB remote sensing estimation method, which realizes accurate monitoring of large-scale evergreen vegetation AGB changes by inverting the VOD index with high spatial resolution and optimizing the evergreen vegetation AGB remote sensing estimation model, so as to provide data support for sustainable forest management, ecological restoration project construction, vegetation ecosystem quality improvement and terrestrial ecosystem carbon cycle research.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] Dataset preparation and preprocessing: collect remote sensing data of the experimental area, including: SMOS satellite passive microwave radiation brightness temperature data, Sentinel-1 satellite active radar backscatter data, SMOS-IC VOD products, AGB spatial reference data, land use data, surface temperature, soil texture and soil moisture data. Based on the time information and spatial resolution in the metadata, select images within the same period, determine the high spatial resolution of the inverted VOD, and perform spatial registration, projection conversion and image cropping.

[0008] The high spatial resolution VOD inversion is coupled with physics and semi-empirical methods. First, the VOD model parameters are extracted. Based on the Sentinel-1 active radar backscattering data, SMOS passive microwave radiation brightness temperature data, surface temperature, soil texture and soil moisture data, the active and passive microwave downscaling model and the hybrid dielectric constant model are used to estimate the high spatial resolution radiation brightness temperature and rough surface effective emissivity parameters respectively. Secondly, the high spatial resolution VOD inversion is combined with the Tau-Omega microwave radiation transmission model to invert the high spatial resolution VOD value of the experimental area. Finally, the accuracy of the VOD value is evaluated. Based on the SMOS-IC VOD product, external interference data and land use data, the accuracy of the inverted high spatial resolution VOD value is evaluated from two aspects: spatial detail performance and interference response capability performance.

[0009] A remote sensing estimation model for AGB of evergreen vegetation was constructed based on the high spatial resolution VOD value and AGB spatial reference data of a certain year. The AGB estimation results at different spatial scales were trained by training the geographic neural network weighted regression model, random forest regression model and XGBoost machine learning model. The model performance test indicators were selected to quantitatively evaluate the estimation effect of each model, and the optimal model and spatial scale for remote sensing estimation of AGB of evergreen vegetation were obtained.

[0010] The spatiotemporal distribution characteristics of evergreen vegetation AGB were extracted. Based on the optimal model of remote sensing estimation of evergreen vegetation AGB, the AGB distribution of evergreen vegetation in the experimental area at the optimal spatial analysis scale was obtained, and the ground AGB reference data of other years were selected for accuracy verification.

[0011] Furthermore, the high spatial resolution VOD inversion specifically includes:

[0012] The radiant brightness temperature with high spatial resolution is generated based on the active and passive microwave downscaling model, and the active and passive microwave downscaling model is expressed as:

[0013]

[0014] where pq represents the cross-polarized radar backscatter observation; CR and FR represent the coarse and fine spatial scales, respectively; θ is the incident angle; the polarization direction of the SMOS satellite p = V; TB p,θ (FR j ) is the radiation brightness temperature of pixel j at the incident angle θ and p polarization direction under FR spatial resolution; TB p,θ (CR) is the radiation brightness temperature of the pixel at the CR spatial resolution when the incident angle is θ and the polarization direction is p; and are the radar backscatter observation values ​​of FR grid pixel j at pp and pq polarizations, respectively. and are the radar backscatter observation values ​​of CR grid pixel j at pp and pq polarizations, respectively; β(CR) represents TB under a single CR grid. p,θ (CR) relative to The sensitivity of the rate of change (in k / dB); Γ(CR) represents the sensitive parameter of the vegetation heterogeneity level within the spatial resolution of CR;

[0015] Based on the mixed dielectric constant model, the effective emissivity parameter of the rough surface is obtained, and the mathematical formula is expressed as follows:

[0016]

[0017] In the formula, r p,θ represents the effective reflectivity of the rough surface; the parameter Q represents the polarization coupling factor; is the Fresnel reflectivity of the smooth surface; ε b is the soil dielectric constant model; the parameter H(SM) represents the effective surface roughness, and the change of H(SM) in the low-frequency L band is measured using the humidity correlation function; NRP p is an angle correction constant, ranging from 0 to 2;

[0018] Based on the high spatial resolution radiation brightness temperature and rough surface effective emissivity parameters, the Tau-Omega microwave radiation transmission model is used to generate high spatial resolution VOD values. The Tau-Omega microwave radiation transmission model expresses the total surface radiation brightness temperature as:

[0019] TB p,θ =T c *(1-ω)*(1-γ)+T c *(1-ω)*(1-γ)*e p,θ *γ+T s *e p,θ *γIn the formula, TB p,θ is the brightness temperature of microwave radiation with an incident angle of θ and in the H-polarization or V-polarization direction; T c and T s represents the real physical temperature of the vegetation canopy and soil surface; ω is the single effective albedo of the surface vegetation; e is the effective emissivity of the rough surface (1-r p,θ ), γ is the vegetation transmittance, and its expression is:

[0020] γ=exp(-VOD / cosθ).

[0021] Furthermore, the VOD accuracy evaluation scheme specifically includes:

[0022] For the spatial detail performance of high spatial resolution VOD, by comparing high spatial resolution VOD with coarse spatial resolution SMOS-IC VOD products, we focus on the differences between the two in different vegetation types, thus highlighting the advantages of high spatial resolution VOD;

[0023] For the disturbance response ability of high spatial resolution VOD, three indicators, namely precipitation, temperature and artificial forest planting area, which have a higher frequency of vegetation disturbance, were selected for evaluation. The evaluation process was divided into three steps. First, the annual change trend of evergreen vegetation in high spatial resolution VOD data was extracted using Z-score (P<0.05) to identify the time when abnormal trend changes occurred; then, the differences between the temperature and precipitation data of the current year and the multi-year average temperature and precipitation values ​​were calculated to obtain the annual temperature and precipitation anomalies; at the same time, the increase or decrease in the area of ​​artificial forests and the change time were counted; finally, these three were compared with the annual change trend of Z-score of high spatial resolution VOD, focusing on comparing whether their change time and trend direction were consistent.

[0024] Preferably, the construction of the evergreen vegetation AGB remote sensing estimation model specifically includes:

[0025] Based on the high spatial resolution VOD index and AGB spatial reference data of a certain year, a geographic neural network weighted regression model, a random forest regression model and an XGBoost machine learning model are selected for training to obtain training results, and the AGB spatial reference data of the certain year is divided into a training set and a test set in a ratio of 7:3;

[0026] Based on the training results of the geographic neural network weighted regression model, random forest regression model and XGBoost machine learning model, the determination coefficient, mean square error and mean absolute error were used to quantitatively evaluate the estimation effect of each model;

[0027] Based on the estimation results of each model, the model and spatial analysis scale with the best performance in estimating the AGB of evergreen vegetation based on VOD were selected.

[0028] Furthermore, the mathematical expression of the geographic neural network weighted regression model is:

[0029] AGB i =W0(μ i ,v i )×β0+W1(μ i ,v i )×β1×VOD i +ε i

[0030] In the formula, i=1,2,…,n; AGB iis the AGB prediction value of the i-th pixel, (μ i ,v i ) is the geographic coordinate of the ith pixel, β0 is the constant term coefficient of the ordinary linear regression model, and β1 is the regression coefficient of the corresponding independent variable in the ordinary linear regression model, reflecting the average level of AGB at the best unbiased estimate. i is the error term for pixel i, with mean zero and variance constant σ 2 .

[0031] Preferably, the mathematical expression of the XGBoost machine learning model is:

[0032]

[0033] In the formula, x i represents the feature parameter of the corresponding pixel in the i-th decision tree, f k (x i ) represents the leaf weight, that is, the regression prediction value of all pixels on the i-th decision tree.

[0034] Preferably, the mathematical expression of the random forest regression model is:

[0035]

[0036] In the formula, ω i is the weight of the base decision tree, usually h i is the predicted value of the base decision tree.

[0037] Furthermore, the evergreen vegetation AGB remote sensing spatiotemporal distribution feature extraction specifically includes:

[0038] Based on the evergreen vegetation AGB remote sensing estimation model, the AGB distribution of evergreen vegetation in all years in the experimental area at the optimal spatial analysis scale was estimated. Using the performance test indicators of the evergreen vegetation AGB remote sensing estimation model, the AGB reference data of the corresponding years were selected to verify the accuracy of the AGB prediction results, and then accurate evergreen vegetation AGB spatiotemporal distribution mapping was obtained.

[0039] Compared with the prior art, the present invention provides a VOD-based evergreen vegetation AGB remote sensing estimation method, which has the following beneficial effects:

[0040] (1) The idea of ​​estimating the AGB of evergreen vegetation based on VOD proposed in this paper not only effectively solves the problem of low data quality and discontinuity caused by the susceptibility of optical remote sensing to weather factors such as clouds, rain, fog and aerosols, but also overcomes the problem that the canopy greenness vegetation index has a spectral saturation effect in densely vegetated areas and cannot effectively capture the changes in AGB in the vertical structure direction of evergreen vegetation.

[0041] (2) The method proposed in this paper combines the strong capture capability of passive microwave satellites for the physiological status of vegetation canopy, vegetation water content and structural changes with the advantages of high spatial resolution and continuous observation of active microwave satellites, thereby improving the spatial resolution of VOD data products and providing a reliable data set for the accurate estimation of AGB of evergreen vegetation.

[0042] (3) The method proposed in the present invention selects spatial details and interference response capabilities to evaluate the data performance of high spatial resolution VOD. This is different from the previous method of only selecting VOD products or NDVI products of the same type and with the same spatial resolution for accuracy evaluation. Instead, it comprehensively considers the degree of presentation of spatial details of ground objects and the consistency of response to environmental interference by high spatial resolution VOD, and can evaluate its data accuracy and performance more objectively and comprehensively.

[0043] (4) The method proposed in the present invention obtains the spatial distribution of evergreen vegetation AGB through an optimized estimation model and spatial analysis scale, reduces the influence of uncertainty factors in the AGB remote sensing estimation process, and can more accurately reflect the spatiotemporal distribution characteristics and development trends of evergreen vegetation AGB, thereby realizing accurate monitoring of evergreen vegetation AGB on a large scale, and providing data support and scientific support for the sustainable development of the ecological environment and the formulation of the dual carbon strategy. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The following drawings will be used to illustrate the specific ideas, structures, and specific processes and results in the embodiments of the present invention in combination, so as to fully reveal the objectives and features of the present invention.

[0045] Figure 1 This is the evergreen vegetation AGB remote sensing estimation process based on VOD in the present invention.

[0046] Figure 2 It is the geographical location of the red soil hilly area in the south of the experimental area of ​​the specific embodiment of the present invention.

[0047] Figure 3 is the SMOS radiant brightness temperature (TB) at a spatial scale of 25 km in the southern red soil hilly area during 2015-2020 25km ) and downscaled to 1 km spatial resolution radiant brightness temperature (TB 1km )result.

[0048] Figure 4 This is the annual spatial distribution of rough surface effective emissivity in the southern red soil hilly area from 2015 to 2020.

[0049] Figure 5 The 1km spatial resolution VOD (VOD) of the southern red soil hilly area during 2015-2020 1km ) spatial distribution each year.

[0050] Figure 6 For the southern red soil hilly area during 2015-2020 VOD 1km Comparison of spatial details with SMOS-IC VOD.

[0051] Figure 7 For the southern red soil hilly area during 2015-2020 VOD 1km Annual change trends of abnormal temperature, abnormal precipitation and afforestation area.

[0052] Figure 8 Parameter test results of AGB estimation for each model under different performance test indicators and spatial analysis scales.

[0053] Fig. 9 This is the spatial distribution of AGB with a spatial resolution of 6 km in the red soil hilly area based on the geographic neural network weighted regression model during 2015-2020.

[0054] Fig.10 The AGB values ​​(AGB) at 6 km spatial resolution in the southern red soil hilly area predicted by the geographic neural network weighted regression model during 2017-2019 Pre_6km ) and AGB spatial reference data (AGB CCI-6km )’s probability density distribution. DETAILED DESCRIPTION

[0055] The technical solutions in the embodiments of the present invention are described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0056] The present invention Figure 1 As shown in the figure, first, the required remote sensing data are collected and preprocessed; then, based on the active and passive microwave remote sensing time series data, the time series regression analysis downscaling model, the hybrid dielectric constant model and the Tau-Omega radiation transmission model are coupled to establish a suitable high spatial resolution VOD inversion model; then, a VOD accuracy evaluation scheme is constructed from two aspects: spatial detail performance and interference response capability; finally, based on the high spatial resolution VOD index and AGB spatial reference data, the model performance test index is used to select the model and spatial analysis scale suitable for the remote sensing estimation of evergreen vegetation AGB, and thereby obtain the spatiotemporal distribution of evergreen vegetation AGB.

[0057] In the embodiment, the southern red soil hilly area is selected as the experimental area of ​​the present invention. The geographical location of the area is as follows: Figure 2 The southern red soil hilly area is the most extensive low mountain and hilly area in my country with a total land area of ​​about 1.24×10 6 km 2The area of ​​soil erosion is about 16×10 4 km 2 The geographical coordinates are 107°-123° east longitude and 18°-34° north latitude, spanning 12 regions including Shanghai, Jiangsu, Zhejiang, Anhui, Fujian, Jiangxi, Henan, Hubei, Hunan, Guangdong, Guangxi and Hainan. The red soil hilly areas are interlaced with evergreen vegetation, with a total cultivated land area of ​​2.82×10 7 hm 2 , carrying the resource needs of nearly 400 million people in my country, plays an important role in ensuring the ecological security of the Yangtze River Economic Belt, regulating the carbon source / carbon sink relationship in the Northern Hemisphere, and maintaining the health and stability of the ecosystem.

[0058] Example

[0059] A VOD-based evergreen vegetation AGB remote sensing estimation method comprises the following steps:

[0060] S1. Multi-source remote sensing data collection and preprocessing: Collect all available SMOS passive microwave radiation brightness temperature data, Sentinel-1 active radar backscatter data, SMOS-IC VOD products, AGB spatial reference data, surface temperature, soil texture and soil moisture data in the experimental area. Based on the time information and spatial resolution in the metadata, screen out images within the same time period, determine the spatial resolution of the inverted VOD data, and perform spatial registration, projection conversion and image cropping.

[0061] In this step, it is necessary to briefly introduce the characteristics of remote sensing data from various sources and the corresponding preprocessing procedures:

[0062] (1) SMOS passive microwave radiation brightness temperature product

[0063] The Level-3 multi-angle radiation brightness temperature products provided by the SMOS satellite are obtained, with a grid size of 25km, a time resolution of 3-7 days, an incidence angle range of 0°-65°, and a data period of 2010-present. This product is a multi-angle dual-polarization / full-polarization atmospheric top-level radiation brightness temperature in orbit units under the antenna polarization reference system and after recalibration. It has the ability to continuously observe all-weather, all-day, and without interference from clouds, rain, fog, and aerosols. For the obtained SMOS Level-3 multi-angle radiation brightness temperature products, firstly, the pixels that are greatly affected by radio frequency interference and solar radiation are removed through the data quality control file; then the selected pixels are converted from non-equidistant projection to equidistant WGS1984 projection, and converted from *.NC to GEOTIFF format; finally, from multiple incidence angles, the angle with the largest number of effective observation pixels is selected, and the image is cropped according to the scope of the experimental area.

[0064] (2) Sentinel-1 active radar backscatter data

[0065] The interferometric wide-band mode ground-range multi-view Level-1 backscatter data from the Sentinel-1 satellite is obtained, with a spatial resolution of 10m, a revisit period of 16 days, and a data period from 2015 to the present. The Sentinel-1 satellite provides free data sets with higher revisit frequency, higher spatial accuracy and wider coverage for long-term series analysis and applications, and also provides key information for the inversion of high-spatial-resolution VOD. For the acquired Sentinel-1 raw data, pre-processing work such as orbit correction, thermal noise removal, radiation calibration, decibel conversion, Doppler terrain correction, incident angle normalization and data clipping is carried out to obtain high-quality Sentinel-1 radar backscatter data.

[0066] (3) AGB spatial reference data

[0067] Get the ESA Biomass CCI annual product produced by the Biomass CCI team, with a spatial resolution of 100m. This product is obtained by combining Earth observation data from Sentinel-1, Envisat's ASAR, JAXA's ALOS-1 / ALOS-2, as well as other field measurement data and space-borne LIDAR data (GEDI and ICESat-2), including the AGB global base map for 2010 and 2017-2020. For the ESA Biomass CCI products obtained, they are mainly used to establish a fitting relationship with high spatial resolution VOD data, and then invert the spatial distribution of evergreen vegetation AGB based on VOD.

[0068] (4) SMOS-IC VOD products

[0069] The SMOS-IC VOD product of version V105 released by the Wigneron team was obtained, with a spatial resolution of 0.25° and a time period of 2010-2020. This product is generated using the minimum cost function iterative inversion algorithm, taking into account the root mean square error between all available multi-orbit radiation brightness temperature observations and simulated radiation brightness temperature of the SMOS satellite and the prior information of all available retrieval parameters. The obtained SMOS-IC VOD product is mainly used to compare and evaluate the spatial detail presentation capability of high spatial resolution VOD data. The preprocessing process of the data includes two main steps: filtering and image cropping. The filtering process uses the root mean square error index between the simulated radiation brightness temperature and the observed radiation brightness temperature to filter out the influence of radio frequency interference.

[0070] (5) Other auxiliary data

[0071] The global annual land cover product with a spatial resolution of 500 m from MCD12Q1 Version 6.12000-2020 was obtained, which was mainly used to extract the evergreen vegetation types in the experimental area that can reflect the dynamic changes of AGB. Data preprocessing included shadow conversion, spatial aggregation, image clipping and mask processing.

[0072] Soil texture data (clay component, organic matter component and sand component) from HarmonizedWorld Soil Database v 1.2 were obtained with a spatial resolution of 0.0083°. This data is mainly used for the inversion of effective reflectance parameters of rough surfaces. Data preprocessing includes spatial connection between database files and raster files, spatial reference definition, raster to point, and point to raster.

[0073] Obtain a seamless daily-scale surface temperature dataset with a global spatial resolution of 1 km from 2003 to 2020 released by Iowa State University, USA. It is mainly used for the VOD inversion of the Tau-Omega radiation transfer model. The data preprocessing process includes projection conversion, Fahrenheit temperature conversion, and image cropping.

[0074] The global 1km spatial resolution daily-scale surface soil moisture dataset from 2000 to 2020 released by the National Tibetan Plateau Science Data Center is obtained. It is mainly used for the VOD inversion of the Tau-Omega radiation transfer model. The data preprocessing process includes projection conversion, scale factor scaling, and image cropping.

[0075] In the present invention, based on the data period and spatial resolution of the above data sources, it is finally determined that the research period for testing the AGB remote sensing estimation method in the experimental area is 2015-2020, and the spatial resolution of the inverted VOD is 1 km.

[0076] S2. High spatial resolution VOD inversion: Based on Sentinel-1 active radar backscattering data, SMOS passive microwave radiation brightness temperature data, soil texture, soil moisture and surface temperature data, the active and passive microwave downscaling model and the hybrid dielectric constant model are used to estimate the high spatial resolution radiation brightness temperature and rough surface effective emissivity parameters, and based on this, combined with the Tau-Omega microwave radiation transmission model, the high spatial resolution VOD distribution of the experimental area is inverted. The specific process is as follows:

[0077] (1) In this step, the time series regression analysis downscaling model establishes a synergistic relationship between active and passive observations through co-polarization and cross-polarization signals. Based on the 10m spatial resolution backscatter information provided by the Senienl-1 satellite, the SMOS radiation brightness temperature at a spatial scale of 25km is downscaled to 1km. The mathematical expression is as follows:

[0078]

[0079] Wherein, pq represents the cross-polarized (VH or HV) radar backscatter observation; CR and FR represent the coarse and fine spatial scales, respectively, corresponding to the 25 km EASE-Grid2.0 grid of the SMOS satellite and the 1 km grid of the Sentinel-1 satellite (generated by aggregating the 10 m grid) described in the present invention; θ is the incident angle, which is defined as 40° in the present invention, which is the angle with the largest number of effective observation pixels for SMOS radiation brightness temperature; the polarization direction of the SMOS satellite p=V; TB p,θ (FR j ) is the SMOS radiation brightness temperature of pixel j at an incident angle of 40° and p-polarization direction at a spatial resolution of 1 km; TB p,θ (CR) is the SMOS radiation brightness temperature of the pixel at an incident angle of 40° and p-polarization direction at a spatial resolution of 25 km; and are the Sentinel-1 radar backscatter observations of 1km grid pixel j at pp and pq polarizations, and are the Sentinel-1 radar backscatter observations of pixel j at 25 km grid in pp and pq polarizations; β(CR) represents TB at a single 25 km grid. p,θ (CR) relative to The sensitivity of the rate of change (in k / dB); Γ(CR) represents the sensitive parameter of vegetation heterogeneity level within a spatial resolution of 25 km, which mainly depends on the sensitivity of radar satellite cross-polarization backscatter observations to vegetation cover and surface roughness.

[0080] SMOS radiant brightness temperature (TB) at 25 km spatial scale in the southern red soil hilly region during 2015-2020 25km ) and downscaled to 1 km spatial resolution radiant brightness temperature (TB 1km ) The result is as follows Figure 3 As shown in the figure, (a) is the annual average TB 1km , (b) is the annual average TB 25km The spatial distribution of TB is shown in Figure 2. (c1)-(c6) is the spatial distribution of TB. 1km and TB 25km The results show that TB 1km and TB 25km The spatial distribution and numerical range of TB are relatively consistent. 25km , TB 1km Covers more sub-pixel spatial detail information of vegetation.

[0081] (2) In this step, the sum of the emissivity of the shallow soil layer and the soil reflectivity is 1. Based on this principle, combined with the high spatial resolution rough surface effective reflectivity inverted by the hybrid dielectric constant model, the effective emissivity information of the rough surface can be obtained. The hybrid dielectric constant model is based on auxiliary information such as soil texture, surface temperature and soil moisture. It converts the microwave radiation and scattering theory of the rough surface into a linear mixture of the Fresnel reflectivity of the smooth surface under different polarization modes, while considering the attenuation effect of the surface roughness on the reflectivity. Finally, the effective emissivity data of the rough surface with a spatial resolution of 1 km in the experimental area is obtained. The mathematical formula is expressed as follows:

[0082]

[0083] In the formula, r p,θ represents the effective reflectivity of the rough surface; the parameter Q represents the polarization coupling factor, which describes the effect of surface roughness on orthogonal polarization; is the Fresnel reflectivity of the smooth surface, generally the soil dielectric constant ε b Function of the incident angle θ (40°); ε b It is the result of the combined action of different components such as soil, air and water. It is necessary to select the corresponding soil dielectric constant model based on the complexity of soil texture. In this invention, considering the physical and chemical properties of the soil in the experimental area, the Mironov mineral soil dielectric constant model and Bircher's organic soil dielectric constant model are introduced to calculate ε b ; The parameter H(SM) represents the effective surface roughness. In the present invention, the change of the low-frequency L-band H(SM) is mainly measured using the piecewise change function of soil moisture; NR p It is an angle correction constant with a value range of 0-2, and is generally used to correct the dependence of surface roughness effect on angle changes.

[0084] The results of the rough surface effective emissivity at a spatial resolution of 1 km in the southern red soil hilly area during 2015-2020 are as follows: Figure 4 As shown in the figure, the results show that the effective emissivity of the rough surface in the experimental area is relatively high, ranging from 0.61 to 0.98. The spatial distribution of the effective emissivity of the surface is relatively consistent every year, and the effective emissivity of the surface in the south is always higher than that in the north.

[0085] (3) In this step, based on the inverted 1km radiation brightness temperature and rough surface effective reflectivity data, the Tau-Omega microwave radiation transmission model is used to generate the VOD with a spatial resolution of 1km (VOD) using the functional relationship between vegetation transmittance and VOD. 1km ) data, the mathematical formula is expressed as:

[0086]

[0087] In the formula, TB p,θ is the brightness temperature of microwave radiation with an incident angle of θ and in the H-polarization or V-polarization direction; T c and T s represents the real physical temperature of the vegetation canopy and soil surface; ω is the single effective albedo of the surface vegetation; γ is the vegetation transmittance, which is usually considered to be a physical quantity related to the vegetation incident angle and the vegetation layer thickness VOD (referring to the moisture thickness from the soil surface to the top of the canopy). The functional relationship between the two is:

[0088] γ=exp(-VOD / cosθ) (4)

[0089] VOD in the red soil hilly areas of southern China during 2015-2020 1km Distribution Figure 5 The results show that the VOD of evergreen vegetation in the red soil hilly area 1km It distributed between 0.12-2.23, with an average value of 2.0±0.21, showing obvious spatial differences.

[0090] S3. Due to the lack of large-scale field site data to verify VOD 1km The accuracy of the inversion is improved by using SMOS-IC VOD products, external interference data and land use data, and the VOD inverted by step S2 is analyzed from two aspects: spatial detail performance and interference response performance. 1km The data were evaluated for accuracy.

[0091] In this step, SMOS-IC VOD products are used to evaluate VOD 1km The spatial detail performance of the two focuses on the differences between them in different vegetation types, especially the number and distribution characteristics of the pixels reflecting different vegetation. Figure 6 As shown in the figure, (a) and (b) are the annual average VOD 1km and the spatial distribution of the annual average SMOS-IC VOD. The results show that VOD 1km It not only presents similar spatial distribution characteristics to SMOS-IC VOD, but also contains more heterogeneous information of evergreen vegetation types.

[0092] On the other hand, three indicators, namely, precipitation, temperature and plantation area, which have a higher frequency of vegetation disturbance, were selected to evaluate VOD. 1km First, the Z-score (P<0.05) was used to extract the VOD 1kmThe annual change trend of evergreen vegetation in the data is used to identify the time when abnormal changes in the trend occur. The annual temperature anomaly, precipitation anomaly, increase or decrease in the area of ​​artificial forests, and the time when they occur are then compared with the VOD 1km The annual change trend of Z-score is compared, focusing on whether their change time and trend direction are consistent, so as to judge the VOD 1km The ability to respond to external interference. 1km The annual trend of temperature anomaly, precipitation anomaly, and plantation area is as follows: Figure 7 As shown in the figure, (a) is VOD 1km (b) The annual variation trend of Z-score (P<0.05) of different vegetation types VOD 1km The Z-score mean, (c) abnormal high temperature, (d) abnormal precipitation, and (e) changes in afforestation area. The results show that VOD 1km It shows strong response ability to abnormal temperature and precipitation disturbances.

[0093] S4. VOD 1km As the core characteristic parameter, based on three representative nonlinear regression models and the AGB spatial reference data of 2020, the AGB estimation results of each model at different spatial scales were obtained through training, and the model performance test indicators were selected to quantitatively evaluate the estimation effect of each model, so as to obtain the optimal model and spatial scale for remote sensing estimation of evergreen vegetation AGB. The specific process is divided into two steps:

[0094] (1) Based on VOD in 2020 1km Based on the AGB spatial reference data, three representative nonlinear regression models and model performance test indicators were selected, and the optimal model parameters for AGB remote sensing estimation of evergreen vegetation were trained from 25 different spatial scales. The 2020 AGB spatial reference data is divided into training set and test set in a ratio of 7:3. The three representative nonlinear regression models are the geographic neural network weighted regression model, the random forest regression model and the XGBoost machine learning model, all of which have outstanding performance in high-dimensional data processing, nonlinear relationship modeling, prediction accuracy, model robustness and anti-fitting ability.

[0095] In this step, the mathematical expression for predicting AGB based on the geographic neural network weighted regression model is:

[0096] AGB i =W0(μ i ,v i )×β0+W1(μ i ,v i )×β1×VOD i+ε i (5)

[0097] In the formula, i=1,2,…,n; AGB i is the AGB prediction value of the i-th pixel, (μ i ,v i ) is the geographic coordinate of the ith pixel, β0 is the constant term coefficient of the ordinary linear regression model, and β1 is the regression coefficient of the corresponding independent variable in the ordinary linear regression model, reflecting the average level of AGB in the entire red soil hilly area at the best unbiased estimation. i is the error term for pixel i, with mean zero and variance constant σ 2 .

[0098] In this step, the mathematical expression for predicting AGB based on the XGBoost machine learning model is:

[0099]

[0100] In the formula, x i represents the feature parameter of the corresponding pixel in the i-th decision tree, f k (x i ) represents the leaf weight, that is, the regression prediction value of all pixels on the i-th decision tree.

[0101] In this step, the mathematical expression for predicting AGB based on the random forest regression model is:

[0102]

[0103] In the formula, ω i is the weight of the base decision tree, usually h i is the predicted value of the base decision tree.

[0104] (2) Based on the training results of the geographic neural network weighted regression model, random forest regression model, and XGBoost machine learning model, the coefficient of determination (R 2 ), mean square error (MSE) and mean absolute error (MAE) and other model performance test indicators are used to quantitatively evaluate the estimation effect of each model at 25 different spatial scales. The calculation formulas of model performance indicators are as follows:

[0105]

[0106] Where n represents the number of pixels; is the AGB prediction value of each model; M i is the AGB observation value of the corresponding pixel, extracted from the pre-divided validation dataset; is the average of the AGB predictions of each model; is the average of the AGB observations. is the standard deviation of the AGB prediction values ​​of each model; M is the standard deviation of the AGB observations. Generally, the smaller the MSE and MAE values ​​are, the higher the reliability of the model prediction results; R 2 The value is between 0 and 1. 2 When it is close to 1, it means that the model parameters have a strong explanatory power for the prediction results.

[0107] (3) Based on the R of each model at different scales 2 , MSE and MAE test results and parameter meanings, by comparing the change rate of each test result at equal intervals, the VOD-based 1km The model and spatial analysis scale that performed best when estimating the AGB of evergreen vegetation in the experimental area. The parameter test results of each model at different scales are shown in Figure 2. Figure 8 As shown in the figure, (a), (b) and (c) are the parameter values ​​of mean square error, mean absolute error and absolute coefficient respectively. The results show that based on VOD 1km The model that performed best when estimating the AGB of evergreen vegetation in the experimental area was the geographic neural network weighted regression model, and the optimal spatial analysis scale was 6 km.

[0108] S5. Extraction of spatiotemporal distribution characteristics of AGB of evergreen vegetation: Based on the geographic neural network weighted regression model constructed and optimized in S4, the AGB time series data of evergreen vegetation at 6 km in the experimental area during 2015-2020 were obtained, and the AGB spatial reference data from 2015 to 2019 were selected for accuracy verification. 2 The spatial distribution of AGB with a spatial resolution of 6 km in the red soil hilly area of ​​southern China based on the geographic neural network weighted regression model during 2015-2020 is shown in Figure 2. Fig. 9 As shown in Figure 2, the AGB estimation results of evergreen vegetation in the experimental area at 6 km and the fitting results of AGB spatial reference data are shown in Figure 2. Fig.10 As shown in the figure, the fitting coefficient between the two is as high as 0.71, indicating that the AGB prediction value of evergreen vegetation in the southern red soil hilly area is relatively reliable.

[0109] 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 remote sensing estimation method for aboveground biomass of evergreen vegetation based on vegetation optical thickness, characterized in that: The method comprises: Dataset preparation and preprocessing: collecting remote sensing data of the experimental area, including: SMOS satellite passive microwave radiation brightness temperature data, Sentinel-1 satellite active radar backscatter data, SMOS-IC vegetation optical thickness products, aboveground biomass spatial reference data, land use data, surface temperature, soil texture and soil moisture data. Based on the time information and spatial resolution in the metadata, images within the same period are screened out to determine the high spatial resolution of the inverted vegetation optical thickness, and spatial registration, projection conversion and image cropping are performed; The high spatial resolution vegetation optical thickness inversion was coupled with physics and semi-empirical methods. First, the parameters of the vegetation optical thickness model were extracted. Based on the Sentinel-1 active radar backscattering data, SMOS passive microwave radiation brightness temperature data, surface temperature, soil texture and soil moisture data, the active and passive microwave downscaling model and the hybrid dielectric constant model were used to estimate the high spatial resolution radiation brightness temperature and rough surface effective emissivity parameters respectively. Secondly, the high spatial resolution vegetation optical thickness inversion was combined with the Tau-Omega microwave radiation transmission model to obtain the high spatial resolution vegetation optical thickness value in the experimental area. Finally, the accuracy of the vegetation optical thickness value was evaluated. Based on the SMOS-IC vegetation optical thickness product, external interference data and land use data, the accuracy of the inverted high spatial resolution vegetation optical thickness data was evaluated from two aspects: spatial detail performance and interference response capability performance. A remote sensing estimation model for aboveground biomass of evergreen vegetation was constructed. Based on the high spatial resolution vegetation optical thickness value and aboveground biomass spatial reference data of a certain year, the remote sensing estimation results of biomass at different spatial scales were trained by using geographic neural network weighted regression model, random forest regression model and XGBoost machine learning model. The model performance test index was selected to quantitatively evaluate the estimation effect of each model, and the optimal model and spatial scale for remote sensing estimation of aboveground biomass of evergreen vegetation were obtained. The spatiotemporal distribution characteristics of the aboveground biomass of evergreen vegetation were extracted. Based on the optimal model for estimating the aboveground biomass of evergreen vegetation, the aboveground biomass distribution of evergreen vegetation in the experimental area at the optimal spatial analysis scale was obtained, and the spatial reference data of the aboveground biomass of the remaining years were selected for accuracy verification.

2. The remote sensing estimation method according to claim 1, characterized in that: The high spatial resolution vegetation optical thickness inversion specifically includes: The radiant brightness temperature with high spatial resolution is generated based on the active and passive microwave downscaling model, and the active and passive microwave downscaling model is expressed as: where pq represents the cross-polarized radar backscatter observation; CR and FR represent the coarse and fine spatial scales, respectively; θ is the incident angle; the polarization direction of the SMOS satellite p = V; TB p,θ (FR j ) is the radiation brightness temperature of pixel j at the incident angle θ and p polarization direction under FR spatial resolution; TB p,θ (CR) is the radiation brightness temperature of the pixel at the CR spatial resolution when the incident angle is θ and the polarization direction is p; and are the radar backscatter observation values ​​of FR grid pixel j at pp and pq polarizations, respectively. and are the radar backscatter observation values ​​of CR grid pixel j at pp and pq polarizations, respectively; β(CR) represents TB under a single CR grid. p,θ (CR) relative to The sensitivity of the change rate; Γ(CR) represents the sensitive parameter of the vegetation heterogeneity level within the spatial resolution of CR; Based on the mixed dielectric constant model, the effective emissivity parameter of the rough surface is obtained, and the mathematical formula is expressed as follows: Where rp,θ represents the effective reflectivity of the rough surface; the parameter Q represents the polarization coupling factor; is the Fresnel reflectivity of the smooth surface; ε b is the soil dielectric constant model; the parameter H(SM) represents the effective surface roughness, and the change of H(SM) in the low-frequency L band is measured using the humidity correlation function; NR p is an angle correction constant, ranging from 0 to 2; Based on the high spatial resolution radiation brightness temperature and rough surface effective emissivity parameters, the Tau-Omega microwave radiation transmission model is used to generate high spatial resolution vegetation optical thickness values. The Tau-Omega microwave radiation transmission model expresses the total surface radiation brightness temperature as: TB p,θ =T c *(1-ω)*(1-γ)+T c *(1-ω)*(1-γ)*e p,θ *c+T s *e p,θ *c In the formula, TB p,θ is the brightness temperature of microwave radiation with an incident angle of θ and in the H-polarization or V-polarization direction; T c and T s represents the real physical temperature of the vegetation canopy and soil surface; ω is the single effective albedo of the surface vegetation; e is the effective emissivity of the rough surface (1-r p,θ ), γ is the vegetation transmittance, and its expression is: γ=exp(-VOD / cosθ) 3. The remote sensing estimation method according to claim 1, characterized in that: The vegetation optical thickness accuracy assessment scheme specifically includes: For the spatial detail expression of high spatial resolution vegetation optical thickness, by comparing high spatial resolution vegetation optical thickness with the coarse spatial resolution SMOS-IC vegetation optical thickness product, we focus on the differences between the two in different vegetation types, thus highlighting the advantages of high spatial resolution vegetation optical thickness; For the disturbance response ability of high spatial resolution vegetation optical depth, three indicators of precipitation, temperature and artificial forest planting area with high vegetation disturbance frequency were selected for evaluation. The evaluation process was divided into three steps. First, the annual change trend of evergreen vegetation in high spatial resolution vegetation optical depth data was extracted using Z-score (P<0.05) to identify the time when its abnormal trend change occurred; then, the differences between the temperature and precipitation data of the current year and the multi-year average temperature and precipitation values ​​were calculated to obtain the annual temperature and precipitation anomalies; at the same time, the increase or decrease in the area of ​​artificial forests and the change time were statistically analyzed; finally, these three were compared with the annual change trend of Z-score of high spatial resolution vegetation optical depth, focusing on comparing whether their change time and trend direction were consistent.

4. The remote sensing estimation method according to claim 1, characterized in that: The construction of the evergreen vegetation aboveground biomass remote sensing estimation model specifically includes: Based on the high spatial resolution vegetation optical thickness index and aboveground biomass spatial reference data of a certain year, a geographic neural network weighted regression model, a random forest regression model and an XGBoost machine learning model were selected for training to obtain training results of different scales. The aboveground biomass spatial reference data of a certain year was divided into a training set and a test set in a ratio of 7:

3. Based on the training results of the geographic neural network weighted regression model, random forest regression model and XGBoost machine learning model, the determination coefficient, mean square error and mean absolute error were used to quantitatively evaluate the estimation effect of each model; Based on the estimation results of each model, the model and spatial analysis scale that performed best in estimating the aboveground biomass of evergreen vegetation based on vegetation optical thickness were selected.

5. The remote sensing estimation method according to claim 4, characterized in that: The mathematical expression of the geographic neural network weighted regression model is: Terms and Conditions i =W0(μ i ,v i )×β0+W1(μ i ,v i )×β1×VOD i +e i In the formula, i=1, 2,...,n; AGB i is the predicted value of aboveground biomass of the ith pixel, (μ i , v i ) is the geographic coordinate of the ith pixel, β0 is the constant term coefficient of the ordinary linear regression model, and β1 is the regression coefficient of the corresponding independent variable in the ordinary linear regression model, reflecting the average level of aboveground biomass at the best unbiased estimate. i is the error term for pixel i, with mean zero and variance constant σ 2 .

6. The remote sensing estimation method according to claim 4, characterized in that: The mathematical expression of the XGBoost machine learning model is: In the formula, x i represents the feature parameter of the corresponding pixel in the i-th decision tree, f k (x i ) represents the leaf weight, that is, the regression prediction value of all pixels on the i-th decision tree.

7. The remote sensing estimation method according to claim 4, characterized in that: The mathematical expression of the random forest regression model is: In the formula, ω i is the weight of the base decision tree, usually h i is the predicted value of the base decision tree.

8. The remote sensing estimation method according to claim 1, characterized in that: The extraction of spatiotemporal distribution characteristics of aboveground biomass of evergreen vegetation specifically includes: Based on the remote sensing estimation model of the aboveground biomass of evergreen vegetation, the aboveground biomass distribution of evergreen vegetation in all years at the optimal spatial analysis scale was estimated in the experimental area. Using the performance test indicators of the remote sensing estimation model of the aboveground biomass of evergreen vegetation, the spatial reference data of the aboveground biomass of the corresponding years were selected to verify the accuracy of the aboveground biomass prediction results of the experimental area, and then accurate spatiotemporal distribution mapping of the aboveground biomass of evergreen vegetation was obtained.

Images