Forest Coverage Extraction Method Based on Hierarchical Multiplication Model and GF-1 Data

Abstract

The present invention relates to a method for extracting forest coverage based on a hierarchical multiplication model and GF-1 data. The present invention includes accurately and efficiently extracting forest coverage from multi-spectral remote sensing data by using a hierarchical multiplication model. On the basis of preprocessing the remote sensing data, first, the land cover types are divided into three main levels according to the differences in the spectral characteristics of the ground objects. Secondly, the abundance of the ground objects is calculated by using a linear pixel decomposition model in each level. Finally, the abundance of forest vegetation in the first-level ground objects is calculated through a multiplication model. The present invention creates a pixel decomposition algorithm based on a multiplication model, improving the applicability and accuracy of extracting the abundance value of ground objects based on multi-spectral images. It solves the limitation of insufficient multi-spectral bands. For the problem that it is difficult to verify the pixel decomposition accuracy, it is proposed to use high-precision images for classification and calculate the abundance of each endmember at a large scale with the classification results to verify the pixel decomposition accuracy, improving the accuracy of pixel decomposition verification.

Description

Technical Field

[0001] The present invention belongs to the technical field of forest remote sensing, and relates to a method for extracting forest coverage based on a hierarchical multiplication model and GF-1 data. Background Art

[0002] Ecological environment protection is a global consensus and a hot issue in today's society. Forestry, as the main body of ecological construction, plays a crucial role in maintaining environmental health and ecological balance, improving the global carbon cycle and water quality, and providing recreational potential. However, factors such as rapid economic growth, rapid population expansion, and frequent natural disasters pose a serious threat to the healthy growth of forestry. Forest coverage rate is an important indicator reflecting the richness of forest resources and the status of ecological balance. Therefore, rapid and efficient forest coverage rate surveys based on high-resolution remote sensing images have become an important research topic in ecological environment construction and protection.

[0003] Traditional monitoring of forest coverage rate and its changes mainly relies on foreign remote sensing data, such as Landsat, MODIS, Spot, Quickbird, etc. However, with the implementation of the major special project of China's high-resolution earth observation system (abbreviated as the high-resolution special project), forest research based on domestic high-resolution satellites has become an important part of forestry remote sensing. Domestic high-resolution data has been widely used in many fields such as estimating forest community area, investigating forest production changes, estimating forest biomass and carbon storage.

[0004] Common methods for extracting forest coverage from remote sensing images include single-band reflectance calculation methods, spectral indices based on algebraic calculations, supervised classification, unsupervised classification, artificial neural networks, expert systems, etc. Among them, the sum of a single spectral band or bands is sensitive to changes in vegetation characteristics and is often used for creating time-series images. The Normalized Difference Moisture Index (NDMI) and Normalized Burn Ratio (NBR) indices are often used to detect forest harvesting and disasters. The Soil Adjusted Vegetation Index (SAVI) and the Atmospherically Resistant Vegetation Index (ARVI) are used to analyze soil effects and atmospheric noise. The accuracy of supervised and unsupervised classification methods is relatively low, and artificial neural networks and expert systems are relatively complex and have poor operability. These methods cannot eliminate the influence of mixed pixels.

[0005] Spectral mixture analysis (SMA) is a mature and effective technique for solving the problem of spectral mixture. SMA models the mixed spectrum as a linear or non-linear combination of spectral endmembers to obtain sub-pixel vegetation information and is widely used in forestry ecological surveys and monitoring. SMA is a method for extracting pure spectral components from mixed pixels, but there are two challenges when applying multi-spectral images such as Gaofen-1 WFV: 1) The number of endmembers is limited by the number of bands in the remote sensing data. According to the theory of mixed pixel decomposition, the maximum number of endmembers is equal to the number of spectral bands plus one. Therefore, it is impossible to accurately extract more land cover types from GF-1 WFV data with only four bands; 2) It is difficult to directly evaluate the unmixing accuracy of mixed pixels. Since it is difficult to determine the true composition of each pixel in the image through ground measurements, direct unmixing accuracy evaluation becomes a difficult problem.

[0006] Therefore, how to provide a method for quickly and accurately extracting forest coverage using multi-spectral remote sensing data with limited bands is an urgent problem to be solved in current forestry remote sensing.

[0007] To solve these problems, a Hierarchical Multiplication Model (HMM) based on pixel decomposition is proposed. The content of this study includes: (1) Using the Hierarchical Multiplication Model (HMM) and GF-1 WFV data to extract the forest coverage of a certain area; 2) Based on the high-resolution (2m) GF-1 PMS data and the SVM method, the land cover types in the experimental area are finely classified, and the forest coverage is calculated according to the classification results, and the accuracy of the HMM results is analyzed using this coverage as verification data. Summary of the Invention

[0008] The object of the present invention is to provide a method for extracting forest coverage based on the Hierarchical Multiplication Model (HMM) of mixed pixel decomposition and GF-1 data, and accurately and efficiently extract forest coverage from multi-spectral remote sensing data using the Hierarchical Multiplication Model. On the basis of preprocessing the remote sensing data, first, the land cover types are divided into three main levels according to the differences in the spectral characteristics of the ground objects. Secondly, the linear pixel decomposition model is used to calculate the abundances of the ground objects in each level, and finally, the abundance of forest vegetation in the first-level ground objects is calculated through the multiplication model.

[0009] The present invention specifically includes the following steps:

[0010] Step 1: Image preprocessing:

[0011] Obtain the GF1 data of the target object, and perform radiometric calibration, atmospheric correction, and geometric correction on the original image by means of the radiometric calibration parameters of the GF1 sensor. Radiometric calibration converts the original DN value into the pixel radiance value. Atmospheric correction uses the FLAASH model (Fast line-of-sight atmospheric analysis of spectral hypercubes) to convert the radiance value into the actual surface reflectance. For geometric correction, first select the homologous ground feature points and perform resampling using the nearest neighbor pixel method, and control the correction error within 0.2 pixels.

[0012] Step 2: Hierarchical classification of land cover types: Classify the land cover types into three hierarchies according to the magnitude of the spectral differences of ground features. The first hierarchy is the water-non-water layer, the second hierarchy is the vegetation-non-vegetation layer, and the third hierarchy is the forest and non-forest layer;

[0013] Step 3: Perform pixel decomposition within each hierarchy to extract the abundances of the corresponding elements in each hierarchy:

[0014] Pixel decomposition is to extract various ground features (endmembers) and the proportions (abundances) of each component from the data of multi-ground feature spectral mixture. The Linear Spectral Mixture Analysis (LSMA) defines that the reflectance of a pixel in a certain spectral band is a linear combination of the reflectances of the basic components that make up the pixel, with the proportion of the pixel area they occupy as the weight coefficient. LSMA is calculated by the following formula:

[0015]

[0016] In the formula, R iλ is the spectral reflectance of pixel i in the λ band; n is the number of ground features; f ki is the proportion of the area occupied by endmember k in pixel i; R kλ is the spectral reflectance of endmember k in the λ band; ε iλ is the error of pixel i in the λ band.

[0017] The proportionality coefficient f k should satisfy the following two constraint conditions:

[0018]

[0019] f ki ≥1 (3);

[0020] The proportionality coefficient f ki is calculated by the least squares method, and the calculation formula is as follows:

[0021]

[0022] where ε iλ , R iλ , r iλ are the error value, observed spectral value, and estimated spectral value of pixel i in the λ band, respectively.

[0023] Select endmembers according to the three levels described in Step 2, and use the LSMA model to perform pixel decomposition at each level to extract the abundances of corresponding ground objects.

[0024] Step 3.1. Extract water bodies and non-water bodies within the first layer:

[0025] The water body endmember is selected from clear water bodies such as rivers, lakes, and reservoirs, and non-water bodies are mainly taken from buildings, roads, vegetation, bare land, etc. Then, use the LSMA model and the least squares method to calculate the abundances of water bodies and non-water bodies. Although the spectral differences between pure water bodies and non-water bodies with low water content are obvious, substances with high water content are greatly affected by moisture. To reduce this interference, the Normalized Difference Water Index (NDWI) is calculated to further correct the extraction errors of water bodies and non-water bodies. The NDWI calculation formula is (green wave - near-infrared) / (green wave + near-infrared). Only non-water body pixels with an NDWI index less than zero can be further decomposed in the next level.

[0026] Step 3.2. Extract vegetation and non-vegetation in the second layer:

[0027] Due to the influence of light, there are significant differences in the spectra of vegetation under light and in the shadow. Therefore, the vegetation in this level is further divided into light vegetation and shadow vegetation, and the vegetation endmembers are selected from vegetation under light conditions and in the shadow, respectively. Non-vegetation is further divided into three categories: bare land, high-light impervious layer, and ordinary impervious layer. The non-vegetation endmembers are selected from the corresponding bare land, high-light buildings, and ordinary impervious layers, respectively. Use the LSMA model to calculate the abundances of light vegetation, shadow vegetation, bare land, high-light impervious layer, and general impervious layer, and then superimpose light vegetation and shadow vegetation as vegetation, and superimpose the remaining subclasses as non-vegetation.

[0028] Step 3.3. Extract forests and non-forests in the third layer:

[0029] Vegetation mainly includes forests, grasslands, and farmlands. The grassland surface is uniform with a high reflectance, the forest surface is rough with a low reflectance, and the reflectance of farmland is between the two. Compared with grasslands and farmlands, forests are significantly affected by light. Therefore, the forests in this layer are first decomposed into light forests and shadow forests, and together with farmlands and grasslands, use the LSMA model to extract their respective abundances. Then, light forests and shadow forests are superimposed as forest types, and farmlands and grasslands are superimposed as non-forest classes.

[0030] Step Four. Calculate the forest coverage using the abundance values and the hierarchical multiplication model (HMM):

[0031] The Hierarchical Multiplication Model (HMM) is a method for calculating abundance problems step by step, with an obvious order of precedence in calculation. The next step is only carried out after the previous step is completed, that is, the abundances of each level are calculated multiple times to obtain the comprehensive abundance. Suppose the first level contains m elements, and the abundances of each element are calculated as f 11 , f 12 , …… f 1m , where each element contains several secondary elements; if the i-th element in the first level contains n secondary elements, its corresponding abundance is f 21 , f 22 , … f 2n , then the abundance of the j-th element in the second level in the first level, that is, the comprehensive abundance, is calculated as f = f 1i * f 2j . And so on, the comprehensive abundance of any element can be obtained by multiplying the abundance of this element at its own level by the comprehensive abundance of the previous level. For example, the formula for calculating the comprehensive abundance of the t-th element at the k-th level is:

[0032] f = f 1i * f 2j * … f kt (5);

[0033] In the process of calculating forest cover, the forest cover calculated by the pixel decomposition model is the forest abundance at the third level, that is, the forest-non-forest level. Its comprehensive abundance is a function of the abundance at the second level of the vegetation type to which it belongs, and the comprehensive abundance of the vegetation is in turn a function of the non-water body abundance at the first level. According to the calculation results of the LSMA model, the forest abundance at the third level is f3, the abundance at the second level of the vegetation type to which the forest belongs is f2, and the abundance at the first level of the non-water body type to which the vegetation belongs is f1. According to the hierarchical multiplication model, the abundance of the forest at the first level, that is, the comprehensive forest cover, is f = f1 * f2 * f3.

[0034] Compared with the prior art, the present invention has the following advantages:

[0035] (1) The present invention creates a pixel decomposition algorithm based on the multiplication model, which improves the applicability and accuracy of extracting the abundance values of ground objects based on multi-spectral images.

[0036] (2) The present invention solves the limitation of insufficient multi-spectral bands and expands the application of multi-spectral images and pixel decomposition algorithms in ground object classification and recognition.

[0037] (3) Aiming at the problem that it is difficult to verify the pixel decomposition accuracy, the present invention proposes to use high-precision images for classification and calculate the abundances of each endmember at a large scale based on the classification results to verify the pixel decomposition accuracy, which improves the accuracy of pixel decomposition verification. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 For the extraction of spectral features of three levels in the embodiment;

[0039] Figure 2 For the forest abundance obtained by the HMM method in the embodiment;

[0040] Figure 3 For the land cover classification results based on SVM and high-resolution images in the embodiment;

[0041] Figure 4 For the forest coverage based on the SVM classification results in the embodiment;

[0042] Figure 5 For the error result distribution of the HMM method in the embodiment;

[0043] Figure 6 For the error distribution analysis in the embodiment. Specific implementation manner

[0044] In order to make the above objects, features and advantages of the present invention more clearly understood, the following will describe the specific implementation manner of the present invention in detail with reference to specific cases.

[0045] In this embodiment, the study area is the land area of Zhejiang Province. The area of Zhejiang is 101,800 square kilometers, with 70.4% mountains, 23.2% plains and basins, and 6.4% water bodies. Most of the mountains are located in the central and southern regions of Zhejiang, with an average altitude of 800m. The plains mainly include the alluvial plains in the north, the coast in the northeast, and the basins in the middle. Zhejiang as a whole shows high in the southwest and low in the northeast. Zhejiang has a distinct subtropical monsoon climate, with four distinct seasons and abundant sunshine. The annual average temperature is 15 to 19 °C, and the average precipitation is 1,100 to 1,900 mm. The superior geographical conditions have given Zhejiang the title of "land of fish and rice".

[0046] The specific steps for extracting the forest coverage of Zhejiang Province based on the hierarchical multiplication model and GF-1 data are as follows:

[0047] Step 1: Download the domestic high-resolution GF-1 wide-field remote sensing image GF-1 WFV from the China Resources Satellite Application Center. The spatial resolution of the image is 16m, and it contains 4 bands. Use the radiometric calibration parameters of the GF1 sensor to perform radiometric calibration on the original image and convert the original DN value into the pixel radiance value; for atmospheric correction, use the FLAASH model to convert the radiance value into the actual surface reflectance to achieve atmospheric correction; select homologous ground control points and use the nearest neighbor pixel method for resampling for geometric correction, and control the correction error within 0.2 pixels.

[0048] Step 2. Stratify land cover types according to the magnitude of spectral differences. The three levels are water - non - water layer, vegetation - non - vegetation layer, and forest and non - forest layer;

[0049] Step 2.1. Select various types such as rivers, lakes, artificial ponds, forests, farmlands, grasslands, bare lands, rocks, urban buildings, roads, etc. from the image, extract the spectral characteristics of each type, and through cluster analysis, determine that the first - level land type classification is water - non - water.

[0050] Step 2.2. Select types such as forest, farmland, grassland, bare land, rock, urban building, road, etc. that are non - water in the first - level classification, extract the spectral characteristics of each type, and through cluster analysis, determine that the second - level land type classification is vegetation and non - vegetation.

[0051] Step 2.3. Select forest, grassland, farmland, etc. that belong to vegetation in the second - level classification and divide them into the third - level types, namely forest and non - forest.

[0052] Step 3. Conduct pixel decomposition within each level and extract the abundances of the corresponding elements in each level.

[0053] Step 3.1. Decompose water and non - water within the first level. Select the end - member information of water and non - water from the remote - sensing image. The water end - members are selected from clear water bodies such as rivers, lakes, and reservoirs, and the non - water is mainly taken from buildings, roads, vegetation, bare lands, etc. As Figure 1 shown in Figure a, the spectral characteristic curves of the water and non - water end - members are significantly different.

[0054] Use the linear pixel decomposition method to extract the abundances of water and non - water. Considering that turbid water bodies and substances with high water content have a high similarity, in order to reduce this interference, the Normalized Difference Water Index (NDWI) is calculated to mask the water bodies. The NDWI calculation formula is (green band - near - infrared band) / (green band+near - infrared band), and only non - water pixels with an NDWI index less than zero can be further decomposed in the next level.

[0055] Step 3.2. Decompose vegetation and non - vegetation within the second level. The surface of the vegetation canopy is rough. Due to the influence of light, there are significant differences in the spectra of vegetation under light and vegetation under shadow. Therefore, the vegetation in this level is further subdivided into light - illuminated vegetation and shadow vegetation, and their end - members are selected from vegetation under light conditions and vegetation under shadow respectively. Non - vegetation includes three types: bare land, highly light - efficient impervious layer, and impervious layer. Their end - members are selected from the corresponding bare land, high - light buildings, and ordinary impervious layers respectively. The characteristic spectral curves of their end - members are as Figure 1 Figure b.

[0056] Based on the above five types of endmembers, the abundances of sunlit vegetation, shaded vegetation, bare land, highly reflective impervious surface, and ordinary impervious surface are obtained using the linear pixel decomposition model. Finally, the abundances of sunlit vegetation and shaded vegetation are combined into the vegetation abundance at this level, and the abundances of the remaining subclasses are combined into the non-vegetation type abundance.

[0057] Step 3.3: Decompose forests and non-forests at the third level. Considering the influence of lighting conditions on the forest spectrum, the forests at this level are first divided into two subclasses: sunlit forests and shaded forests. The endmembers of sunlit forests are selected from the top of the canopy and areas with better lighting conditions, while the endmembers of shaded forests are selected from the backlit canopy. Non-forests mainly include farmland and grassland, and their endmembers are selected from farmland and grassland areas respectively. The spectral characteristic curves of their endmembers are shown in Figure 1 c,

[0058] Using the above endmember spectra and the linear pixel decomposition method, the abundances of sunlit forests, shaded forests, farmland, and grassland at this level are extracted. Finally, the abundances of sunlit forests and shaded forests are merged into the forest abundance at this level, and the abundances of farmland and grassland are merged into the non-forest abundance at this level.

[0059] Step Four: Calculate the forest coverage using the abundance values and the multiplication model.

[0060] The comprehensive forest coverage in the study area is a function of the forest abundance at the third level, the vegetation abundance at the second level, and the non-water body abundance at the first level. In the third-level forest-non-forest hierarchy, the forest abundance extracted by LSMA is f forest , and the forest belongs to the vegetation type at the second level. The vegetation abundance is calculated by the LSMA model as f vegetation , while the vegetation belongs to the non-water body type at the first level. The non-water body abundance is calculated by the LSMA model as f non-water , and is calculated using the multiplication model. Therefore, the comprehensive forest coverage in the study area is calculated using the multiplication model, and the calculation formula is: f = f non-water *, f vegetation * f forest , and the calculation results are shown in Figure 2 as shown.

[0061] Step Five: Test the accuracy of the HMM method. The accuracy test is mainly achieved through two methods:

[0062] a) Calculate the accuracy of the HMM method using the official statistical data as the verification data:

[0063] Obtain the forest coverage of Zhejiang Province through statistical data, and use this data as verification data. Compare the forest coverage of the study area obtained by the HMM method with the verification data to analyze the overall accuracy and calculation error of the HMM method. The statistical report of the Forestry Bureau of Zhejiang Province shows that the forest coverage of Zhejiang in 2019 was 61.15%, while the comprehensive forest coverage calculated by the HMM method was 60.09%, and the calculation error was only 1.57%.

[0064] b) Calculate the vegetation coverage using high-spatial-resolution data as verification data to analyze the accuracy of the HMM method.

[0065] Select a 1.5 km * 1.2 km area in the study area that contains all land cover types in this study as the test area. Download the 2-meter-resolution GF-1 panchromatic / multispectral remote sensing data (GF-1PMS) covering the test area. Use the support vector machine to classify the surface of the verification area, and calculate the forest coverage of the test area based on the classification results as verification data.

[0066] ① Perform radiometric correction, atmospheric correction, and geometric correction on the GF-1PMS image, and perform multispectral and panchromatic fusion to obtain a 2-meter-resolution image. In ENVI, start Toolbox / Image Sharpening / NNDiffuse PanSharpening, select the orthorectified multispectral and panchromatic images respectively, select the output path and file name, and the fused image is a 2m-resolution multispectral image.

[0067] ② Use the support vector machine for supervised classification based on the fused image. The classification types are the same as those of the HMM classification, including water bodies, forests, bare land, high-reflection objects, ordinary buildings, farmland, grasslands, etc. Open the PMS fused image in ENVI, create corresponding ROI files for 8 classes according to the number of image classifications. Double-click the newly created ROI file with the left mouse button, a tool box will pop up, modify the ROI name, select the representative color, select the corresponding area on the image, and draw an irregular closed polygon to generate ROI samples of water bodies, forests, bare land, high-reflection objects, ordinary buildings, farmland, and grasslands, and save them. In the software toolbar Toolbox, select "Raster Management-Masking-Building Mask", select the PMS fused image to be masked, and create a mask file. In the software toolbar Toolbox, click "Classification-Supervised Classification" and select "Support Vector Machine Classification". In the pop-up dialog box, select the fused image, mask file, and classification samples, and output the file name to complete the supervised classification, such as Figure 3As shown, all ground objects are classified into two major categories: forest and non-forest according to the classification results.

[0068] ③ Based on the supervised classification results, calculate the forest abundance in the verification area. In the test area, grids of 16 * 16 meters are divided according to the resolution of WFV. Each grid contains 8 * 8 = 64 pixels with a spatial resolution of 2m. Figure 4 As shown, since the type of each small pixel has been identified as forest or non-forest, the proportion of the number of small pixels classified as forest type in the entire large pixel is calculated, which is the fine forest abundance (coverage rate) of the test area.

[0069] ④ Analyze the accuracy of the HMM method using the forest abundance calculated from the fused image. By comparing the forest abundance obtained by the method combining high-precision image and SVM with the HMM method, it can be found that their correlation reaches 0.95. In terms of spatial distribution, the correlation inside the forest is high. As Figure 5 shown, the result errors are mainly distributed in the forest edge areas.

[0070] As Figure 6 shown, through statistics, it is found that the forest coverage of 43.12% of the pixels is exactly the same, the proportion of pixels with an error within 0.1 is 25.30%, the pixels with an error within 0.2 are 15.9, the proportion of pixels with an error within 0.3 is 9.08%, the proportion of pixels with an error within 0.4 is 1.46%, and only 2.61% of the pixels have an error exceeding 0.4.

[0071] The forest coverage extraction method based on the hierarchical multiplication model and GF-1 data provided by the present invention makes full use of the pixel decomposition model to reduce the adverse effects of mixed pixels in the forest coverage extraction process, and at the same time uses the multiplicative transfer model to solve the problem of insufficient bands in the multi-spectral image during the pixel decomposition process. It not only greatly improves the accuracy of forest coverage extraction, but also expands the application of multi-spectral images in forest coverage extraction. The application of domestic GF-1 remote sensing data provides strong data support for large-scale forest coverage extraction.

[0072] It can also combine time series data to accurately calculate the forestry coverage, monitor forest disturbance information, and analyze the factors affecting forest ecological changes.

Claims

1. A method for extracting forest coverage based on a hierarchical multiplication model and GF-1 data, characterized in that: The specific steps include: Step 1: Image preprocessing: Obtain the Gaofen-1 GF1 data of the target object, and use the radiation calibration parameters of the GF1 sensor to perform radiation calibration, atmospheric correction, and geometric correction on the original image; Step 2: Land cover type stratification: According to the size of the spectrum difference of the ground objects, it is divided into the first layer of water body-non-water body layer, the second layer of vegetation-non-vegetation layer, and the third layer of forest and non-forest layer; Step 3: Decompose pixels in each level and extract the abundance of corresponding elements in each level: Step 3.1: Decompose water and non-water bodies in the first layer: Water body end members and non-water body end members were selected, and the abundance of water body and non-water body was calculated using the linear spectral mixture model (LSMA) model and the least squares method. The interference of high water content substances was reduced by calculating the normalized water index (NDWI). The NDWI calculation formula is: (green wave - near infrared) / (green wave + near infrared). Only non-water body pixels with an NDWI index less than zero can be further decomposed in the next level. Step 3.2 Decompose vegetation and non-vegetation in the second layer: Vegetation is divided into illuminated vegetation and shaded vegetation, and vegetation end members are selected from vegetation under illuminated conditions and vegetation under shaded conditions respectively; non-vegetation is divided into three categories: bare land, impervious layer with high light efficiency and impervious layer, and non-vegetation end members are selected from the corresponding bare land, high-light building and ordinary impervious layer respectively; the LSMA model is used to obtain the abundance of illuminated vegetation, shaded vegetation, bare land, high-light impervious layer and ordinary impervious layer respectively, and then the illuminated vegetation and shaded vegetation are superimposed as vegetation, and the remaining subclasses are superimposed as non-vegetation; Step 3.3: Decompose forest and non-forest in the third layer: Forest endmembers are divided into illuminated forest and shadowed forest, and their abundances are extracted from non-forest endmembers using the LSMA model. Then, illuminated forest and shadowed forest are superimposed as forest types, and the rest are superimposed as non-forest types. Step 4: Calculate forest coverage using abundance values ​​and the hierarchical multiplicative model HMM: Forest cover is a function of the abundance of forests at the lowest level and the abundance of its type at the next level above, recursively to the highest level; According to the calculation results of the LSMA model, the forest abundance at the third level is f3, the abundance of the vegetation type to which the forest belongs at the second level is f2, and the abundance of the non-water body type to which the vegetation belongs at the first level is f1. According to the hierarchical multiplication model, the calculation formula for the abundance of the forest at the first level, i.e., the forest coverage, is: f=f1*f2*f3.

2. The forest coverage extraction method based on the hierarchical multiplication model and GF-1 data according to claim 1, wherein: The radiation calibration described in step 1 converts the original DN value into the pixel radiation brightness value. The atmospheric correction uses the FLAASH model to convert the radiation brightness value into the actual surface reflectance. The geometric correction first selects the ground object points with the same name and resamples them using the nearest neighbor pixel method. The correction error is controlled within 0.2 pixels.

3. The forest coverage extraction method based on the hierarchical multiplication model and GF-1 data according to claim 1, characterized in that: The pixel decomposition described in step 3 is to extract various objects and the proportion of each component from the data of multi-object spectral mixture. The linear spectral mixture model LSMA defines that the reflectance of a pixel in a certain spectral band is a linear combination of the reflectance of the basic components constituting the pixel and the proportion of the pixel area occupied by them as weight coefficients; The LSMA calculation method is: ； ； ； Where is the spectral reflectance of pixel i in the λ band; n is the number of ground objects; is the area proportion of endmember k within pixel i; is the spectral reflectance of endmember k in the λ band; is the error of pixel i in the λ band; area proportion is calculated by the least squares method, and the calculation formula is as follows: ； where , are the error, spectral reflectance, and estimated spectral value of pixel i in the λ band, respectively; Select endmembers according to the three levels in Step 2 respectively, and use the LSMA model to perform pixel decomposition at each level to extract the abundances of corresponding ground objects.

Images