Fine inversion method and system for forest and grass vegetation coverage based on multi-scale remote sensing

Abstract

The application discloses a kind of based on multi-scale remote sensing's forest grass vegetation coverage refinement inversion method and system, first provide end member type and end member abundance information using unmanned aerial vehicle field observation data, then these prior knowledge is input linear mixed pixel decomposition model and is obtained corresponding end member spectral information of end member type, with the aid of unmanned aerial vehicle more than resolution observation data as prior knowledge, improve the end member spectral extraction precision, provide technical support for improving forest grass vegetation total coverage inversion precision;Make full use of the texture features of high-resolution satellite observation data, combine the spectral characteristics of medium resolution with the texture features of high resolution, use random forest model to construct the best response relationship between different remote sensing feature combination and woody vegetation coverage measured data, so as to realize the high-precision inversion of large-area woody vegetation coverage, provide data support for further estimating herbaceous vegetation coverage.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing image processing, and particularly relates to a fine inversion method and system for forest and grass vegetation coverage based on multi-scale remote sensing. BACKGROUND

[0002] Vegetation coverage refers to the proportion of the area of vegetation vertical projection to the ground in unit area. Vegetation coverage is an important basic data and key parameter in the research of surface processes such as ecological hydrology and soil erosion process, so the precision and quality of vegetation coverage data have always been concerned and valued by many experts and scholars. Traditional vegetation distribution investigation is mainly based on ground observation, which has certain limitations and costs a lot of manpower and material resources. There is also great uncertainty in the extrapolation based on field measurement sample plots. With the continuous development of remote sensing observation technology, it is possible to remotely sense vegetation coverage in a large area, and the research on vegetation coverage inversion based on remote sensing data has always been a hot spot in the field of quantitative remote sensing of vegetation.

[0003] At present, the vegetation coverage inversion methods based on remote sensing technology mainly include empirical model, physical model and mixed pixel decomposition model. The empirical model mainly establishes the statistical relationship between the remote sensing features such as vegetation index or band reflectance and the measured vegetation coverage to invert the regional vegetation coverage; the physical model mainly quantitatively inverts the vegetation canopy leaf area index from the remote sensing observation ground reflectance signal by using the vegetation radiation transfer mechanism model, and then converts the leaf area index into canopy vegetation coverage by using the canopy porosity formula; the mixed pixel decomposition method is also widely used in the estimation of vegetation coverage in semi-arid areas, which uses the pixel decomposition model to decompose the brightness information contribution of each end member in the pixel to obtain the end member vegetation abundance to estimate the coverage. In arid / semi-arid areas, sparse woody vegetation, herbaceous vegetation and dry vegetation are the main components of forest and grass vegetation, but the widely used vegetation coverage inversion algorithm cannot finely identify the coverage of woody vegetation, herbaceous vegetation and dry vegetation, which seriously affects the inversion accuracy of the total coverage of forest and grass vegetation.

[0004] In addition, the existing MODIS and Landsat tree cover remote sensing products are directly constructed by machine learning algorithms to build the relationship model between the measured sample and the low-resolution remote sensing spectral features, and are applied to large areas. However, the inversion accuracy of this method is low in the sparse wood vegetation, herbaceous vegetation and dry vegetation mixed scene area in the arid and semiarid region, because the spectral features of herbaceous vegetation and wood vegetation are very similar, and the spectral features of dry vegetation and bare land are very similar, which seriously affects the inversion accuracy. After verifying the published Tree Cover product with the measured tree cover data in the semiarid region, it is found that the existing Tree Cover product seriously underestimates the tree cover in the semiarid region, which cannot meet the application requirements. At present, there are few studies on the fine remote sensing inversion of forest and grass vegetation coverage under the mixed and complex terrain of various land cover types. The few studies reported mainly use the "two-step method" to separate wood vegetation and herbaceous vegetation, which is based on high-resolution remote sensing images to identify wood vegetation canopy and then calculate the coverage, but the inversion accuracy of this method is affected by the classification accuracy of wood vegetation canopy, and it is not suitable for large-scale regional expansion application. The mixed pixel decomposition model is used to separate green vegetation, dry vegetation and bare land, but the inversion accuracy of this method is affected by the endmember spectrum accuracy of green vegetation, dry vegetation and bare land. Therefore, it is urgent to propose a practical fine remote sensing inversion method for forest and grass vegetation coverage in a large area.

[0005] In summary, the current vegetation coverage inversion method based on remote sensing technology mainly has the following defects:

[0006] (1) Empirical model method refers to directly establishing the response relationship model between vegetation index or band reflectance and measured vegetation coverage, and using the model to extrapolate the vegetation coverage in a large area. The mathematical method for constructing the response relationship generally uses statistical regression and machine learning algorithm. However, when this method is applied to the inversion of tree cover, the mixed reflection signal of wood vegetation, herbaceous vegetation, dry vegetation and bare land is included in the low-resolution remote sensing observation spectrum, which cannot obtain the fine coverage of different components in forest and grass vegetation, and the empirical model parameters cannot be used.

[0007] (2) The physical model method also quantitatively inverts the leaf area index of vegetation canopy through spectral reflectance, but the low-resolution remote sensing observation spectrum includes the reflectance of wood vegetation and herbaceous vegetation, so it can only quantitatively invert the total leaf area index of forest and grass vegetation through the vegetation radiation transfer model, and cannot separate the leaf area index of wood vegetation and herbaceous vegetation, so it cannot separate the coverage of wood vegetation and herbaceous vegetation through the canopy porosity formula. Therefore, this method can only invert the total coverage of forest and grass vegetation, and also does not consider the dry vegetation coverage.

[0008] (3) The principle of the mixed pixel decomposition model method is that one pixel in the image may actually be composed of multiple components, each of which contributes to the information observed by the remote sensing sensor, so that the remote sensing information can be decomposed to establish a pixel decomposition model, and the vegetation coverage can be estimated by using the model. The pixel binary model is the simplest model in the linear pixel decomposition model method. The pixel binary model assumes that a pixel is composed of only two parts, vegetation-covered ground and non-vegetation-covered ground, and the mixed spectral information is only linearly synthesized by the two component factors, and the ratio of the area of each factor in the pixel is the weight of each factor. The percentage of the vegetation-covered ground in the pixel is the vegetation coverage of the pixel, which is expressed by the formula:

[0009] F cover =(NDVI-NDVI min ) / (NDVI max -NDVI min )

[0010] In the formula, NDVI min is the minimum NDVI value of the non-vegetation-covered ground, and NDVI max is the maximum NDVI value of the vegetation-covered ground.

[0011] However, in the mixed area of woody vegetation and herbaceous vegetation, the spectral reflection characteristics of herbaceous and woody vegetation in the visible and infrared bands are similar, and the NDVI value is very close. This method cannot accurately separate woody vegetation and herbaceous vegetation, and does not consider the influence of dry vegetation on the total coverage of forest and grass vegetation. Although part of the mixed pixel decomposition model considers the end member of dry vegetation and green vegetation, in a high heterogeneity area where multiple vegetation coverage types are mixed and the vegetation site conditions are complex, there may be multiple end member spectra for the same type of vegetation. It is difficult to directly find high-quality end member spectra of different vegetation types on medium and low resolution remote sensing images, so it cannot support the mixed pixel decomposition model to estimate the abundance of green vegetation, dry vegetation and soil components with high precision, and it is also impossible to decompose the woody and herbaceous vegetation coverage. Therefore, an automatic method for obtaining green vegetation, dry vegetation and bare soil multi-end member spectra is urgently needed to improve the accuracy of end member spectrum extraction and thus improve the accuracy level of the mixed pixel decomposition model. SUMMARY

[0012] In order to solve the problem that the existing method is low in inversion of green vegetation and dry vegetation coverage, and cannot separate the coverage of woody vegetation and herbaceous vegetation in green vegetation, the present application provides a fine inversion method and system for forest and grass vegetation coverage based on multi-scale remote sensing, which uses the idea of hierarchical classification inversion to perform layer-by-layer inversion on the coverage of woody vegetation, herbaceous vegetation and dry vegetation. It can be applied to fine inversion of forest and grass vegetation coverage (woody vegetation coverage, herbaceous coverage and dry vegetation coverage) in arid / semi-arid areas, and can provide new technical means for high-precision monitoring of vegetation dynamic process, and can also provide high-precision underlying surface data support for evaluation of vegetation ecological hydrological benefit based on ecological model, hydrological model and soil erosion model and other surface water and soil process models.

[0013] In order to achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0014] The present application provides a fine inversion method for forest and grass vegetation coverage based on multi-scale remote sensing, comprising:

[0015] Step 1: obtaining ground sample plot unmanned aerial vehicle image, and classifying the obtained unmanned aerial vehicle image into green vegetation, dry vegetation and soil;

[0016] Step 2: taking the average value of the surface reflectivity of the Landsat pixel corresponding to the ground sample plot range coordinates as the mixed pixel observation signal value of the sample Landsat; taking green vegetation, dry vegetation and soil as the endmember type within the pixel; based on the classification result of each sample unmanned aerial vehicle image, obtaining the coverage value of each endmember type, and taking the coverage value as the prior knowledge of the abundance of the endmember type;

[0017] Step 3: based on the prior knowledge of the abundance of the endmember type, using a linear mixed pixel decomposition model to inversely calculate the endmember spectrum to obtain an endmember spectrum library;

[0018] Step 4: based on the obtained endmember spectrum library data, using a multi-endmember mixed pixel decomposition model to forwardly invert the abundance of green vegetation, dry vegetation and soil to obtain the abundance of green vegetation, dry vegetation and soil at a regional scale, and then obtaining the dry vegetation coverage, and calculating the sum of the abundance of green vegetation and dry vegetation to obtain the forest and grass vegetation coverage;

[0019] Step 5: using a man-machine interactive interpretation method to obtain woody vegetation in each sample unmanned aerial vehicle image, and calculating the woody vegetation coverage to construct a woody vegetation coverage sample data set;

[0020] Step 6: extracting the spectral features and texture features of Landsat and high-resolution satellite image data;

[0021] Step 7: The sample data set of woody vegetation coverage is taken as a target variable, the extracted spectral features, texture features and the forest and grass vegetation coverage obtained in step 4 are taken as characteristic variables, and a random forest machine learning model is used to build a nonlinear relationship between the woody vegetation coverage and the spectral features, texture features and the forest and grass vegetation coverage, so as to predict the woody vegetation coverage;

[0022] Step 8: According to the physical shielding relationship between the forest and grass vegetation coverage and the woody vegetation coverage and the herbaceous vegetation coverage, the herbaceous vegetation coverage is inversely calculated.

[0023] Further, in step 1, the obtained unmanned aerial vehicle image is classified in the following manner:

[0024] Step 1.1: Based on the field investigation photos and the unmanned aerial vehicle image, green vegetation, dry vegetation and soil samples are selected;

[0025] Step 1.2: The RGB color values of the unmanned aerial vehicle image are converted by using a color space change function, so as to obtain 12-dimension color features of the 4 color spaces of RGB, HSV, XYZ and L * a * b * ;

[0026] Step 1.3: The color features after the RGB conversion are automatically classified by using a support vector machine supervised classification method, and the unmanned aerial vehicle image classification results of the green vegetation, dry vegetation and soil are obtained by combining a visual interpretation method.

[0027] Further, the calculation formula of the linear mixed pixel decomposition model is as follows:

[0028] x k =M T f k +ε

[0029] Wherein, x k is a b×1 vector, representing the ground reflectivity of the kth sample, and b is the number of bands; f k is a c×1 vector, representing the end member coverage of the kth sample, and c is the number of end member categories; M is a c×b matrix, representing the end member spectrum matrix, and each row array element represents the spectrum of a specific end member type; and ε represents the model fitting error.

[0030] Further, in step 3, for n samples, the linear mixed pixel decomposition model is modified as follows:

[0031] X=FM

[0032] In the formula, X is an n×b matrix, X=[x1,x2,…,x k ,…,x n ]T ; F is an nxc matrix, F = [f1, f2, … fk k ,…f n ] T ;

[0033] The endmember spectrum matrix M of the study area is obtained by using the least square method:

[0034] M = (F T F) -1 F T X.

[0035] Further, the calculation formula of the multi-endmember mixed pixel decomposition model is:

[0036]

[0037] In the formula, X is the actual observation value of the image pixel; N is the number of endmember categories; d i is the number of endmember spectra of the ith endmember type; e ij is the jth endmember spectrum of the ith endmember; p ij is the proportion of the jth spectrum of the ith endmember in the pixel; q is 0 or 1, representing the label of whether the endmember spectrum participates in the unmixing; n is the error component; 0≤p ij ≤1 and

[0038] Further, in step 4, the multi-endmember mixed pixel decomposition model with the smallest root mean square error of the pixel spectrum simulation value in the calculation process is selected as the optimal spectral unmixing model, and the p ij value obtained by the optimal spectral unmixing model is the abundance value of each endmember type in the pixel. The abundance of green vegetation and dry vegetation calculated by the optimal spectral unmixing model is the green vegetation coverage.

[0039] Further, the spectral features include all-band surface reflectance, normalized vegetation index, visible band difference vegetation index, normalized difference water index, ratio vegetation index, enhanced vegetation index, modified soil adjusted vegetation index, difference vegetation index, green normalized vegetation index, green band index and near-infrared vegetation index.

[0040] Further, in step 6, the texture features of all-band surface reflectance are calculated by using the gray level co-occurrence matrix based on statistical analysis.

[0041] Further, in step 8, the herbaceous vegetation coverage is obtained by back calculation according to the following formula:

[0042] f total = f woody +(1-fwoody ) x f herb

[0043] wherein f total is the forest vegetation coverage, f woody and f herb are the coverage of woody vegetation and herbaceous vegetation respectively.

[0044] Another aspect of the present application provides a multi-scale remote sensing-based forest and grass vegetation coverage refinement inversion system, comprising:

[0045] A ground sample plot unmanned aerial vehicle image acquisition and classification module is configured to acquire ground sample plot unmanned aerial vehicle images and classify the acquired unmanned aerial vehicle images into green vegetation, dry vegetation and soil.

[0046] An end member type coverage value derivation module is configured to take the average value of the surface reflectivity of the Landsat pixel corresponding to the ground sample plot range coordinates as the mixed pixel observation signal value of the Landsat of the sample plot; take the green vegetation, dry vegetation and soil as the end member types within the pixel; based on the classification results of each sample plot unmanned aerial vehicle image, obtain the coverage values of each end member type, and take the coverage values as the prior knowledge of the abundance of the end member types.

[0047] An end member spectral library inversion module is configured to, based on the prior knowledge of the abundance of the end member types, utilize a linear mixed pixel decomposition model to perform reverse calculation on the end member spectrum to obtain an end member spectral library.

[0048] A multi-end member mixed pixel decomposition module is configured to, based on the obtained end member spectral library data, utilize a multi-end member mixed pixel decomposition model to perform forward inversion on the abundance of the green vegetation, dry vegetation and soil to obtain the abundance of the green vegetation, dry vegetation and soil at a regional scale, and further obtain the dry vegetation coverage and calculate the sum of the abundance of the green vegetation and dry vegetation to obtain the forest and grass vegetation coverage.

[0049] A woody vegetation coverage sample data set construction module is configured to acquire the woody vegetation in each sample plot unmanned aerial vehicle image by using a human-computer interactive interpretation method and calculate the woody vegetation coverage to thereby construct a woody vegetation coverage sample data set.

[0050] A feature variable set construction module is configured to extract the spectral features and texture features of the Landsat and high-resolution satellite image data.

[0051] The woody vegetation coverage prediction module is configured to take the woody vegetation coverage sample data set as a target variable, take the extracted spectral features, the texture features and the forest and grass vegetation coverage obtained by the multi-end member mixed pixel decomposition module as feature variables, and use a random forest machine learning model to construct a nonlinear relationship between the woody vegetation coverage and the spectral features, the texture features and the forest and grass vegetation coverage, so as to predict the woody vegetation coverage.

[0052] The herbaceous vegetation coverage prediction module is configured to inversely calculate the herbaceous vegetation coverage according to a physical shielding relationship between the forest and grass vegetation coverage and the woody vegetation coverage and the herbaceous vegetation coverage.

[0053] Compared with the prior art, the present application has the beneficial effects that:

[0054] The existing mixed pixel decomposition is mainly to extract end member spectra directly on remote sensing images, but in the sparse forest and grassland area in the arid / semi-arid region, it is difficult to effectively find end member spectra of 100% abundance end member type, thereby affecting the analysis accuracy of the mixed pixel decomposition model. The present application uses the unmanned aerial vehicle field observation data to provide prior knowledge of end member type and end member abundance, uses a multi-end member mixed pixel spectral decomposition model to inversely calculate the end member spectrum theoretical value, and can improve the accuracy of extracting the end member spectrum.

[0055] The existing woody vegetation coverage inversion is mainly to use the spectral features of medium and low resolution satellite remote sensing to inversely calculate the woody vegetation coverage, but the inversion effect in the sparse forest and grassland area in the arid / semi-arid region is poor, because the spectral features of the herbaceous vegetation are very similar to those of the woody vegetation, and the herbaceous vegetation seriously affects the inversion accuracy of the woody vegetation coverage. The present application fully utilizes the texture features of high resolution satellite (remote sensing) image data to reflect the difference between the woody vegetation and the herbaceous vegetation, combines the medium resolution spectral features and the high resolution texture features, uses a random forest model to construct the best response relationship between different remote sensing feature combinations and the measured data of the woody vegetation coverage, can improve the inversion accuracy of the woody vegetation coverage, and provides data support for the estimation of the herbaceous vegetation coverage. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 A flowchart of a fine inversion method of forest and grass vegetation coverage based on multi-scale remote sensing is provided for the embodiment of the present application;

[0057] Figure 2 A composition structure diagram of an end member spectrum library is provided for the embodiment of the present application;

[0058] Figure 3 A multi-end member mixed pixel decomposition flowchart is provided for the embodiment of the present application;

[0059] Figure 4A schematic diagram of an architecture of a forest and grass vegetation coverage fine inversion system based on multi-scale remote sensing is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0060] The present application will be further explained in conjunction with the accompanying drawings and specific embodiments:

[0061] As shown in the drawings, Figure 1 a forest and grass vegetation coverage fine inversion method based on multi-scale remote sensing comprises:

[0062] S101: Obtain ground sample plot unmanned aerial vehicle images, and classify the obtained unmanned aerial vehicle images into green vegetation, dry vegetation and soil;

[0063] S102: Take the average value of the surface reflectivity of the Landsat pixel corresponding to the ground sample plot range coordinates as the mixed pixel observation signal value of the Landsat of the sample plot; take the green vegetation, dry vegetation and soil as the endmember types within the pixel; based on the classification results of each sample plot unmanned aerial vehicle image, obtain the coverage values of each endmember type, and take the coverage values as the abundance prior knowledge of the endmember types;

[0064] S103: Based on the abundance prior knowledge of the endmember types, use a linear mixed pixel decomposition model to perform reverse calculation on the endmember spectrum to obtain an endmember spectrum library;

[0065] S104: Based on the obtained endmember spectrum library data, use a multi-endmember mixed pixel decomposition model to perform forward inversion on the abundance of the green vegetation, dry vegetation and soil to obtain the abundance of the green vegetation, dry vegetation and soil at the regional scale, and further obtain the dry vegetation coverage, and calculate the sum of the abundance of the green vegetation and dry vegetation to obtain the forest and grass vegetation coverage;

[0066] S105: Obtain the woody vegetation in each sample plot unmanned aerial vehicle image using a man-machine interactive interpretation method, and calculate the woody vegetation coverage to construct a woody vegetation coverage sample data set;

[0067] S106: Extract the spectral features and texture features of the Landsat and high-resolution satellite image data;

[0068] S107: Take the woody vegetation coverage sample data set as the target variable, take the extracted spectral features, texture features and the forest and grass vegetation coverage obtained in S104 as the characteristic variables, use a random forest machine learning model to construct the non-linear relationship between the woody vegetation coverage and the spectral features, texture features and forest and grass vegetation coverage, and predict the woody vegetation coverage;

[0069] S108: According to the physical shielding relationship between the forest and grass vegetation coverage and the woody vegetation coverage, the herbaceous vegetation coverage is obtained by back calculation.

[0070] The method specifically comprises:

[0071] 1. Total forest and grass vegetation coverage inversion

[0072] (1) Ground sample unmanned aerial vehicle image acquisition and classification

[0073] In the target area, a 30m*30m sample is carried out, and the center and four corners of the sample are accurately positioned by using differential RTK GPS. An unmanned aerial vehicle is used to vertically photograph each sample, so that the overlapping degree of each picture is more than 60%, the shooting range can completely cover the 30m*30m ground sample, and professional software is used to splice and ortho-correct the unmanned aerial vehicle pictures. It is ensured that the spatial distribution of woody vegetation, herbaceous vegetation, dry vegetation and soil in the unmanned aerial vehicle photograph can be clearly distinguished.

[0074] The present application designs an unmanned aerial vehicle image automatic classification processing flow: ① Based on the field investigation photos and unmanned aerial vehicle images, green vegetation, dry vegetation and soil samples are selected; ② The RGB color value of the unmanned aerial vehicle image is converted by using the color space change function, and 12-dimensional color features of 4 color spaces (RGB, HSV, XYZ and L * a * b * ) are obtained, and the calculation formula is shown in (1)-(5); ③ The color features after RGB conversion are automatically classified by using support vector machine supervised classification method, and the high-precision unmanned aerial vehicle image classification results of green vegetation, dry vegetation and soil are obtained by combining visual interpretation method.

[0075] The calculation formula of converting RGB to HSV is as follows:

[0076]

[0077] In the formula:

[0078]

[0079] The calculation formula of converting RGB to XYZ is as follows:

[0080]

[0081] The calculation formula of converting RGB to L * a * b * is as follows:

[0082]

[0083] where:

[0084]

[0085] where X n , Y n and Z n are the values for a particular white object point.

[0086] (2) Endmember spectral library inversion

[0087] The average value of the surface reflectance in the 2x2 area adjacent to the Landsat pixel corresponding to the plot range coordinates is taken as the mixed pixel observation signal value of the Landsat remote sensing satellite for the plot; green vegetation, dry vegetation and soil are taken as the endmember types within the pixel; based on the classification results of the unmanned aerial vehicle image for each plot (including green vegetation, dry vegetation and soil), the coverage value of a type can be calculated by dividing the total number of pixels of the type by the total number of all pixels in the plot, and this coverage value is taken as the prior knowledge of the abundance of the endmember type.

[0088] The linear mixed pixel decomposition model (LSMA) is used to calculate the endmember spectrum in reverse, and the calculation formula is as follows:

[0089] x k = M T f k + ε (6)

[0090] In the formula, x k is a b x 1 vector representing the surface reflectance of the kth plot, and b is the number of bands; f k is a c x 1 vector representing the endmember coverage in the kth plot, and c is the number of endmember categories; M is a c x b matrix, and each row array element represents the spectrum of a specific endmember type; M matrix is called endmember spectrum matrix; ε represents the model fitting error.

[0091] For n plots, the linear mixed pixel decomposition model can be modified as:

[0092] X = FM (7)

[0093] In the formula, X is an n x b matrix, X = [x1, x2, …, x k ,…, x n ] T ; F is an n x c matrix, F = [f1, f2, … f k ,… f n ] T .

[0094] The least squares method can be used to solve the endmember spectrum matrix M of the study area:

[0095] M = (FT F) -1 F T X(8)

[0096] (3) Multi-terminal hybrid pixel decomposition

[0097] The endmember spectral library consists of three classes: A, B, and C. Class A endmembers include different spectra such as A1, A2, A3, etc. (metaspectral diversity), class B endmembers include B1, etc., and so on. When constructing the unmixing model, one or more spectra, or none at all, are selected from each endmember class as the model's endmember spectra for unmixing. The structure of the endmember spectral library is as follows: Figure 2 .

[0098] Based on the endmember spectral library data obtained through the above steps, the abundance of the three components—green vegetation, dry vegetation, and soil—is forward inverted using the Multi-Endmember Mixed Pixel Decomposition Model (MESMA). The calculation process is as follows: Figure 3 As shown. The MESMA calculation formula is as follows:

[0099]

[0100] In the formula, X represents the actual observed value of the image pixel; N represents the number of endmember categories; d i e is the number of endmember spectra of the i-th type of endmember; ij p is the spectrum of the j-th endmember of the i-th type of endmember; ij Let be the proportion of the j-th spectrum of the i-th endmember in the pixel; q is 0 or 1, representing a label indicating whether this endmember spectrum participates in the unmixing; n is the error component. Similar to the linear mixing model, the multi-endmember mixing pixel decomposition model also has two constraints: 0 ≤ p ij ≤1 and

[0101] The root mean square error (RMSE) of the simulated pixel spectrum values ​​is selected during the calculation process. s The smallest MESMA model is used as the optimal model for spectral unmixing, and the p calculated by this model ij This refers to the abundance value of each endmember type within that pixel.

[0102]

[0103] Where m represents the number of spectral bands, ε b This represents the root mean square error of the simulated pixel value for band b.

[0104] The model (i.e. the optimal model of spectral unmixing) can be used to calculate the abundance of green vegetation, dry vegetation and soil at the regional scale, wherein the green vegetation abundance represents the coverage of green woody and green herbaceous vegetation, the dry vegetation abundance represents the coverage of yellow herbaceous vegetation (dry woody vegetation does not produce coverage because it has no canopy), i.e. the dry vegetation coverage. Therefore, the sum of the green vegetation and dry vegetation abundance is the forest and grass vegetation coverage.

[0105] 2 Fine inversion of woody vegetation and herbaceous vegetation coverage

[0106] (1) Construction of woody vegetation coverage sample data set

[0107] Due to the large uncertainty of the automatic classification results of green woody vegetation and green herbaceous vegetation, in order to improve the accuracy of the sample data, the present application adopts a human-computer interactive interpretation method to obtain the woody vegetation canopy in each quadrat unmanned aerial image and calculate the woody vegetation coverage, thereby constructing a high-quality sample data set of woody vegetation coverage.

[0108] (2) Construction of feature variable set

[0109] Spectral features and texture features of Landsat (visible, near-infrared, shortwave infrared) and high-resolution satellite image data (visible, near-infrared) were extracted. The spectral features included all-band surface reflectance, Normalized Difference Vegetation Index (NDVI), Visible-band Difference Vegetation Index (VDVI), Normalized Difference Moisture Index (NDMI), Ratio Vegetation index (RVI), Enhanced Vegetation Index (EVI), Modified Soil-Adjusted Vegetation Index (MSAVI), Difference Vegetation Index (DVI), Green Normalized Difference Vegetation Index (GNDVI), Green-Red Vegetation index (GRVI) and Near-Infrared Reflectance of Vegetation (NIRv). The texture features of all-band surface reflectance were calculated by Gray-Level-Co-occurrence Matrix (GLCM) based on statistical analysis.

[0110] (3) Woody vegetation cover prediction model construction

[0111] The sample data set of woody vegetation cover was taken as the target variable, and the spectral features, texture features of Landsat and high-resolution satellite image data corresponding to the sample area and the grass vegetation cover inversion result in step 1 were taken as the characteristic variables. The random forest machine learning model was used to construct the complex nonlinear relationship between the woody vegetation cover and the spectral features, texture features and grass vegetation cover, and to construct the remote sensing inversion model of the woody vegetation cover at the regional scale. Then, the recursive feature elimination (RFE) was used to perform optimal variable selection on all the prediction variables, so as to determine the most robust and highest-precision machine learning prediction model of the woody vegetation cover.

[0112] (4) Herbaceous vegetation cover prediction model construction

[0113] According to the physical shielding relationship between the total coverage of vegetation (i.e. the coverage of forest and grass vegetation) and the coverage of woody vegetation and the coverage of herbaceous vegetation in the vertical projection direction, the coverage of herbaceous vegetation can be obtained by using formula (11).

[0114] f total = f woody (1-f woody ) x f herb (11)

[0115] In the formula, f total is the coverage of forest and grass vegetation, f woody and f herb are the coverage of woody vegetation and the coverage of herbaceous vegetation, respectively.

[0116] On the basis of the above-mentioned embodiments, as shown in the formula (11), the present application further provides a fine inversion system for the coverage of forest and grass vegetation based on multi-scale remote sensing, comprising: Figure 4

[0117] The ground sample unmanned aerial vehicle image acquisition and classification module is used for acquiring the ground sample unmanned aerial vehicle image and classifying the acquired unmanned aerial vehicle image into green vegetation, dry vegetation and soil.

[0118] The end member type coverage value derivation module is used for taking the average value of the surface reflectivity of the Landsat pixel corresponding to the ground sample range coordinates as the mixed pixel observation signal value of the sample Landsat, taking the green vegetation, the dry vegetation and the soil as the end member types in the pixel, obtaining the coverage values of each end member type based on the classification results of each sample unmanned aerial vehicle image, and taking the coverage values as the prior knowledge of the abundance of the end member types.

[0119] The end member spectral library inversion module is used for inversely calculating the end member spectrum based on the prior knowledge of the abundance of the end member types by using a linear mixed pixel decomposition model, so as to obtain an end member spectral library.

[0120] The multi-end member mixed pixel decomposition module is used for forward inversion of the abundance of the green vegetation, the dry vegetation and the soil by using a multi-end member mixed pixel decomposition model based on the acquired end member spectral library data, so as to obtain the abundance of the green vegetation, the dry vegetation and the soil at a regional scale, further obtain the coverage of dry vegetation, and calculate the sum of the abundance of the green vegetation and the dry vegetation to obtain the coverage of forest and grass vegetation.

[0121] The woody vegetation coverage sample data set construction module is used for acquiring the woody vegetation in each sample unmanned aerial vehicle image by using a man-machine interactive interpretation mode, calculating the coverage of woody vegetation, and thus constructing a woody vegetation coverage sample data set.

[0122] ​A feature variable set construction module is configured to extract spectral features and texture features of Landsat and high-resolution satellite image data;

[0123] A woody vegetation coverage prediction module is configured to use a random forest machine learning model to construct a nonlinear relationship between the woody vegetation coverage and the spectral features, the texture features and the forest-grass vegetation coverage, and to predict the woody vegetation coverage by taking the forest-grass vegetation coverage as a target variable and the extracted spectral features, texture features and the forest-grass vegetation coverage obtained by the multi-end member mixed pixel decomposition module as feature variables.

[0124] A herbaceous vegetation coverage prediction module is configured to inversely calculate the herbaceous vegetation coverage according to a physical shielding relationship among the forest-grass vegetation coverage, the woody vegetation coverage and the herbaceous vegetation coverage.

[0125] In summary, due to the strong spatial heterogeneity of the surface landscape under complex terrain conditions, it is difficult to directly obtain end member spectra from satellite remote sensing images in actual applications, and the method for indirectly extracting end member spectra only uses coarse-scale observation information of medium and low resolution satellites, and it is difficult to effectively extract end member types and spectral information that can meet the real coverage of the ground. To solve this problem, the present application innovatively uses unmanned aerial vehicle field observation data to first provide end member type and end member abundance information, and then inputs this prior knowledge into a linear mixed pixel decomposition model to inversely calculate the end member spectral information of the corresponding end member type. This idea improves the traditional method of directly selecting end member spectra from remote sensing images, and improves the end member spectral extraction accuracy by using the observation data of the unmanned aerial vehicle as prior knowledge, thereby providing technical support for improving the forest-grass vegetation total coverage inversion accuracy.

[0126] Due to the similar spectral features of woody and herbaceous vegetation, the traditional vegetation coverage inversion method (empirical model, mechanism model or mixed pixel decomposition) based on medium resolution spectral features cannot finely separate the woody and herbaceous vegetation coverage, and thus there is a large error in the inversion result of the woody vegetation coverage. To solve this problem, the present application fully utilizes the texture features of high-resolution satellite observation data, combines the medium resolution spectral features and the high resolution texture features, and uses a random forest model to construct the best response relationship between different remote sensing feature combinations and the measured data of the woody vegetation coverage, so as to realize high-precision inversion of the woody vegetation coverage in a large area, thereby providing data support for the estimation of the herbaceous vegetation coverage.

[0127] The above only shows the preferred embodiments of the present application, and it should be noted that those skilled in the art can make several improvements and refinements without departing from the principles of the present application, and these improvements and refinements should also be considered as the protection scope of the present application.

Claims

1. A multi-scale remote sensing-based fine inversion method for forest and grass vegetation coverage, characterized in that, The method comprises the following steps: Step 1: acquiring ground sample plot unmanned aerial vehicle images, and classifying the acquired unmanned aerial vehicle images into green vegetation, dry vegetation and soil; Step 2: taking the average value of the ground reflectance of the Landsat pixel corresponding to the ground sample plot range coordinate as the mixed pixel observation signal value of the Landsat of the sample plot; taking the green vegetation, dry vegetation and soil as the endmember types in the pixel; based on the classification result of each sample plot unmanned aerial vehicle image, obtaining the coverage values of the endmember types, and taking the coverage values as the prior knowledge of the abundance of the endmember types; Step 3: based on the prior knowledge of the abundance of the endmember types, the endmember spectrum is reversely calculated by using a linear mixed pixel decomposition model to obtain an endmember spectrum library; Step 4: based on the acquired endmember spectrum library data, the abundance of the green vegetation, dry vegetation and soil is forwardly inversed by using a multi-endmember mixed pixel decomposition model to obtain the abundance of the green vegetation, dry vegetation and soil at a regional scale, and then the dry vegetation coverage is obtained, and the sum of the abundance of the green vegetation and the dry vegetation is taken as the forest and grass vegetation coverage; Step 5: the wood vegetation in each sample plot unmanned aerial vehicle image is acquired by using a man-machine interactive interpretation mode, and the wood vegetation coverage is calculated to construct a wood vegetation coverage sample data set; Step 6: extracting spectral features and texture features of Landsat and high-resolution satellite image data; Step 7: taking the wood vegetation coverage sample data set as a target variable, taking the extracted spectral features, texture features and the forest and grass vegetation coverage obtained in step 4 as characteristic variables, and using a random forest machine learning model to construct a nonlinear relationship between the wood vegetation coverage and the spectral features, texture features and forest and grass vegetation coverage, and the wood vegetation coverage is predicted; Step 8: according to the physical shielding relationship among the forest and grass vegetation coverage, the wood vegetation coverage and the herbaceous vegetation coverage, the herbaceous vegetation coverage is inversely calculated; The calculation formula of the multi-endmember mixed pixel decomposition model is: In the formula, X is the actual observation value of the image pixel; N is the number of endmember categories; d i is the number of endmember spectra of the ith endmember category; e ij is the jth endmember spectrum of the ith endmember category; p ij is the proportion of the jth spectrum of the ith endmember in the pixel; q is 0 or 1, representing the label of whether the endmember spectrum participates in unmixing; n is the error component; 0 ij ≤ 1 and In step 4, the multi-end member mixed pixel decomposition model with the minimum root mean square error of the simulated value of the pixel spectrum in the calculation process is selected as the optimal spectral unmixing model, and p ij That is, the abundance value of each end member type in the pixel, and the abundance of green vegetation, dry vegetation and soil at the regional scale is calculated by using the optimal spectral unmixing model, wherein the sum of the abundance of green vegetation and dry vegetation is the forest and grass vegetation coverage. In step 6, the texture features of the ground reflectance of all bands are calculated by using a gray level co-occurrence matrix based on statistical analysis.

2. The method according to claim 1, wherein, In step 1, the acquired unmanned aerial vehicle images are classified in the following manner: Step 1.1: selecting green vegetation, dry vegetation and soil samples based on field investigation photos and unmanned aerial vehicle images; Step 1.2: Transform the RGB color values of the UAV image using the color space change function to obtain the color features of 12 dimensions of RGB, HSV, XYZ and L * a * b * ​ Step 1.3: using a support vector machine supervised classification method to automatically classify the color features after RGB transformation, and combining a visual interpretation method to acquire the unmanned aerial vehicle image classification results of the green vegetation, dry vegetation and soil.

3. The method according to claim 1, wherein, The calculation formula of the linear mixed pixel decomposition model is: x k = M T f k + ε where x k is a b × 1 vector representing the surface reflectance of the kth sample, b is the number of bands; f k is a c × 1 vector representing the endmember coverage in the kth sample, c is the number of endmember classes; M is a c × b matrix representing the endmember spectral matrix, each row of which contains the spectral elements of a particular endmember type; and ε represents the model fitting error.

4. The method according to claim 3, wherein, In step 3, for n sample plots, the linear mixed pixel decomposition model is modified as: X = FM In the formula, X is an n×b matrix, X=[x1,x2,…,x k ,…,x n ] T F is an n×c matrix, F = [f1, f2, ..., fc] k ,…f n ] T ; The endmember spectrum matrix M of the research area is solved by using a least square method: M = (F T F) -1 F T X.

5. The method according to claim 1, wherein, The spectral features include the ground reflectance of all bands, normalized vegetation index, visible band difference vegetation index, normalized difference water index, ratio vegetation index, enhanced vegetation index, modified soil adjusted vegetation index, difference vegetation index, green normalized vegetation index, green band index and near-infrared vegetation index.

6. The method according to claim 1, wherein, In step 8, the herbaceous vegetation coverage is obtained according to the following formula: f total = f woody + (1 - f woody ) x f herb wherein f total is the forest and grass vegetation coverage, f woody and f herb are the coverage of woody vegetation and herbaceous vegetation, respectively.

7. A multi-scale remote sensing-based forest and grass vegetation coverage refinement inversion system, characterized in that, The application comprises the following steps: The ground sample unmanned aerial vehicle image acquisition and classification module is used for acquiring the ground sample unmanned aerial vehicle image and classifying the acquired unmanned aerial vehicle image into green vegetation, dry vegetation and soil; The end member type coverage value derivation module is used for taking the average value of the ground surface reflectivity of the Landsat pixel corresponding to the ground sample range coordinates as the mixed pixel observation signal value of the sample Landsat; taking the green vegetation, dry vegetation and soil as the end member types in the pixel; based on the classification result of each sample unmanned aerial vehicle image, the coverage values of the end member types are obtained, and the coverage values are taken as the prior knowledge of the abundance of the end member types; The end member spectral library inversion module is used for inversely calculating the end member spectrum based on the prior knowledge of the abundance of the end member types by using a linear mixed pixel decomposition model to obtain an end member spectral library; The multi-end member mixed pixel decomposition module is used for forward inversion of the abundance of the green vegetation, dry vegetation and soil by using a multi-end member mixed pixel decomposition model based on the acquired end member spectral library data, to obtain the abundance of the green vegetation, dry vegetation and soil at a regional scale, and then obtain the dry vegetation coverage and calculate the sum of the abundance of the green vegetation and the dry vegetation to obtain the forest and grass vegetation coverage; The woody vegetation coverage sample data set construction module is used for acquiring the woody vegetation in each sample unmanned aerial vehicle image by using a man-machine interactive interpretation mode and calculating the woody vegetation coverage, so as to construct a woody vegetation coverage sample data set; The feature variable set construction module is used for extracting the spectral features and texture features of the Landsat and high-resolution satellite image data; The woody vegetation coverage prediction module is used for taking the woody vegetation coverage sample data set as a target variable, taking the extracted spectral features, texture features and forest and grass vegetation coverage obtained by the multi-end member mixed pixel decomposition module as feature variables, and constructing a nonlinear relationship between the woody vegetation coverage and the spectral features, texture features and forest and grass vegetation coverage by using a random forest machine learning model, to predict the woody vegetation coverage; The herbaceous vegetation coverage prediction module is used for inversely calculating the herbaceous vegetation coverage according to the physical shielding relationship among the forest and grass vegetation coverage, the woody vegetation coverage and the herbaceous vegetation coverage; The calculation formula of the multi-end member mixed pixel decomposition model is as follows: In the formula, X is the actual observation value of the image pixel; N is the number of endmember categories; d i is the number of endmember spectra of the ith endmember category; e ij is the jth endmember spectrum of the ith endmember category; p ij is the proportion of the jth endmember spectrum of the ith endmember category in the pixel; q is 0 or 1, representing the label of whether the endmember spectrum participates in unmixing; n is the error component; 0 ij ≤ 1 and In the multi-end member mixed pixel decomposition module, a multi-end member mixed pixel decomposition model with the minimum root mean square error of the simulated spectral value of the end member is selected as an optimal spectral unmixing model in the calculation process, and p ij That is, the abundance value of each end member type in the pixel, and the abundance of green vegetation, dry vegetation and soil at the regional scale is calculated by using the optimal spectral unmixing model, wherein the sum of the abundance of the green vegetation and the dry vegetation is the forest and grass vegetation coverage. In the feature variable set construction module, the texture features of the ground surface reflectivity of all bands are calculated by using a gray level co-occurrence matrix based on statistical analysis.

Images