Method for retrieving vertical structure characteristics of forest canopy, method for classifying large-area forest and system

Abstract

The application provides a forest canopy vertical structure feature inversion method, a large-area forest classification method and system, and belongs to the field of forest ecological environment monitoring. The inversion method collects two different year data in a long time sequence, respectively infers a canopy height spatial distribution map, extracts a pixel with increased canopy height as a height sample set, and extracts the rest as a feature sample set, to obtain forest canopy height data; then, spectral features and spatial texture features are obtained; the year when forest disturbance occurs is taken as a growth starting point, and the canopy height, spectral reflectance and spatial texture of each pixel from the growth starting point to the last year are obtained; the average growth speed feature of the last year or when the canopy spectral reflectance reaches saturation is extracted, and the canopy height feature and the growth speed feature are used to represent the canopy vertical structure feature; and the classification method fuses horizontal space-time spectrum features and vertical structure features to perform forest classification. The application improves the accuracy and precision of large-area forest classification.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of forest ecological environment monitoring, and particularly relates to a forest canopy vertical structure feature inversion method, a large-area forest classification method and system. BACKGROUND

[0002] Forests are an important part of the terrestrial ecosystem, and play a key role in preventing soil erosion, maintaining biodiversity, maintaining global carbon-oxygen balance, and mitigating global warming. However, different forest types have significant differences in regional soil and water conservation, watershed water and heat circulation, and local climate regulation, and their ecological benefits and carbon sink capacity are also different. For example, the photosynthesis, soil and water conservation, and nutrient conservation capacity of broad-leaved forests are better than those of coniferous forests, the carbon sequestration capacity of coniferous forests is higher than that of broad-leaved forests, but the ecological environmental stability of coniferous forests is slightly inferior to that of broad-leaved forests, and the soil nutrient content, forest temperature and humidity, vegetation coverage, and water and soil loss prevention capacity of mixed forests are all better than those of pure forests. Therefore, large-area, high-precision, and fine forest type spatial distribution has become the basic data for modern forest resource optimization, sustainable management, and ecological system service evaluation, and high-precision forest type classification method research has become a research hotspot in forest resource monitoring and sustainable management.

[0003] Forest classification generally includes spectral remote sensing, laser radar, GPS measurement, area sampling, and other methods. Among them, spectral remote sensing data can be obtained by satellite, aerial, unmanned aerial vehicle, and other means. Satellite remote sensing technology has become an important technical means for regional, national, and global forest type remote sensing classification due to its wide coverage, short revisit period, and low data acquisition cost.

[0004] In the prior art, according to the difference of classification features used, the forest type remote sensing classification method based on satellite remote sensing data classifies forests from the aspects of canopy spectral response characteristic difference, special feature difference, canopy spatiotemporal spectral feature difference, and canopy vertical structure features relying on airborne LiDAR data. However, the above classification methods have problems such as insufficient classification precision, insufficient comprehensiveness, insufficient classification accuracy, and small classification range, and cannot meet the needs of fine forest resource monitoring and sustainable management. SUMMARY

[0005] In view of the above defects or deficiencies in the prior art, the present application aims to provide a forest canopy vertical structure feature inversion method, a large-area forest classification method and system provided by an embodiment of the present application, which combines canopy horizontal spatiotemporal spectral features and vertical structure features to improve the accuracy and precision of classification.

[0006] In order to achieve the above-mentioned purpose, the embodiments of the present application adopt the following technical solutions:

[0007] In a first aspect, the embodiments of the present application provide a forest canopy vertical structure feature inversion method, comprising the following steps:

[0008] In step S101, a monitoring area is determined, sample point data and long-time optical remote sensing data of two different years are collected, and a forest canopy height spatial distribution map of each year is inferred;

[0009] In step S102, according to the canopy height spatial distribution map, a pixel with increased forest canopy height in the second year than in the first year is extracted as a height sample set, and a pixel without increased forest canopy height is extracted as a feature sample set;

[0010] In step S103, the height difference and the feature difference corresponding to the second year and the first year are calculated respectively, a relationship model of the height difference and the feature difference is constructed, and forest canopy height data in the long time sequence is obtained;

[0011] In step S104, remote sensing images in the long time sequence of the monitoring area are obtained, and after preprocessing, normalized difference vegetation index (NDVI) is used to represent the forest canopy spectral feature, and spectral variation vegetation index variance (SVVI_svar) is used to represent the spatial texture feature;

[0012] In step S105, missing values are filled to obtain a complete forest canopy height data set, an NDVI set and an SVVI_svar set;

[0013] In step S106, forest disturbance detection data in the long time sequence of the monitoring area are obtained, and a year in which forest disturbance occurs is taken as a growth starting point;

[0014] In step S107, based on the forest canopy height data set, the NDVI set and the SVVI_svar set, canopy height, spectral reflectance and spatial texture of each pixel from the growth starting point to the last year in the long time sequence are obtained, and the growth trajectory of the forest type is quantitatively described;

[0015] In step S108, from the year in which forest disturbance occurs to the last year in the long time sequence or when the forest canopy spectral reflectance reaches saturation, the forest canopy spectral feature, the spatial texture feature and the canopy height feature of each pixel are extracted to represent the forest canopy growth speed feature;

[0016] In step S109, the forest canopy height feature and the growth speed feature are used to represent the forest canopy vertical structure feature.

[0017] As a preferred embodiment of the present application, in step S103, the height difference and the feature difference corresponding to the second year and the first year are calculated according to the height sample set and the feature sample set respectively.

[0018] In a second aspect, the embodiments of the present application also provide a large-area forest classification method, which comprises:

[0019] Step S1, determine the monitoring area and the long time interval, and obtain multi-source remote sensing data of the monitoring area;

[0020] Step S2, extract the forest canopy level spatio-temporal spectral features using the multi-source remote sensing data;

[0021] Step S3, obtain the vertical structure features between forest types by using the forest canopy vertical structure feature inversion method as described above.

[0022] Step S4, fuse the forest canopy level spatio-temporal spectral features and the vertical structure features, and use the forest type sample obtained by field investigation and the machine learning method to perform large-scale forest type remote sensing classification;

[0023] Step S5, quantitatively evaluate the mapping accuracy by using the forest field investigation sample data of the target year.

[0024] As a preferred embodiment of the present application, the multi-source remote sensing data in step S1 includes SRTM DEM data, Landsat series data in the long time interval, spaceborne LiDAR data, Sentinel-1 / 2 data, meteorological data, soil data, and forest type field investigation data.

[0025] As a preferred embodiment of the present application, step S4 specifically includes:

[0026] Step S41, fuse the extracted forest canopy level spatio-temporal spectral features, forest canopy height features, and growth speed features to jointly constitute a feature space for forest type remote sensing classification;

[0027] Step S42: based on the forest canopy level spatio-temporal spectral features and the vertical structure features, calculate the feature correlation and feature importance by using the association hierarchical clustering method and the random forest algorithm RF to select classification features beneficial to forest type remote sensing classification in the monitoring area, and collect all the beneficial classification features as an optimal classification feature combination;

[0028] Step S43, based on the optimal classification feature combination, perform forest type remote sensing classification by using the RF classifier.

[0029] As a preferred embodiment of the present application, the forest canopy level spatio-temporal spectral features extracted in step S3 include spectrum, phenology, texture, polarization, and geographic environment.

[0030] As a preferred embodiment of the present application, when extracting the texture features, 18 texture features calculated by using the RedEdge1 band and the gray level co-occurrence matrix are used to comprehensively represent the spatial texture feature differences of the forest vegetation canopy.

[0031] As a preferred embodiment of the present application, when extracting the phenological feature, 14 statistical values of the time series data within a year are selected; the statistical values include: 10-day interval red edge position index REP, comprehensive statistical features of the time series data, intra-annual vertical polarization VV, vertical-horizontal polarization VH, standard deviation of the modified vegetation index mRVI, intra-annual and winter normalized difference vegetation index maximum difference NDVI_maxsummer, and summer and winter mRVI maximum difference mRVI_summerwinter; wherein, for the extraction of the comprehensive statistical features of the time series data, the linear interpolation method and the Savitzky-Golay filtering algorithm are used to generate the REP time series data set at 10-day intervals within a year, and the mean, standard deviation, maximum value, minimum value, median, range, coefficient of variation, 0.25 quantile, 0.75 quantile, interquartile range, interquartile range, and interquartile range are calculated using all the REP time series data to represent the phenological feature differences of different forest types.

[0032] As a preferred embodiment of the present application, when extracting the geographical environment feature, three topographic features of elevation, slope and aspect are extracted; four climate features of annual mean temperature, seasonal mean temperature, annual mean precipitation and seasonal mean precipitation are extracted based on global 1km resolution, monthly scale climate products within a long time interval; the soil moisture response index SMRI is calculated using the water loss rate model based on the 1km resolution, daily scale, 10-100cm depth soil moisture data product within a year, which is used to quantify the ability of soil to provide water required for the growth and development of forest types; and the bilinear interpolation method is used to resample all geographical environment covariates to 30m spatial resolution.

[0033] In a third aspect, the embodiments of the present application further provide a large-area forest classification system, which comprises: a data acquisition module, a horizontal spatio-temporal spectral feature extraction module, a vertical structure feature extraction module, a remote sensing classification module and an evaluation module; wherein,

[0034] The data acquisition module is used to determine a monitoring area and a long time interval, and acquire multi-source remote sensing data of the monitoring area;

[0035] The horizontal spatio-temporal spectral feature extraction module is used to extract the forest canopy horizontal spatio-temporal spectral feature using the multi-source remote sensing data;

[0036] The vertical structure feature extraction module is used to acquire the vertical structure feature between forest types by using the forest canopy vertical structure feature inversion method as described above;

[0037] The remote sensing classification module is used to fuse the forest canopy horizontal spatio-temporal spectral feature and the vertical structure feature, and perform large-scale forest type remote sensing classification using the forest type samples obtained by field investigation and the machine learning method;

[0038] The evaluation module is used for quantitatively evaluating the mapping accuracy by using the forest field investigation sample data of the target year.

[0039] The technical solution provided by the embodiment of the application has the following beneficial effects:

[0040] The forest canopy vertical structure feature inversion method, the large-area forest classification method and the system provided by the embodiment of the application improve the classification effect of large-area forest types, improve the accuracy and precision of classification, and solve the core pain point of high cost in large-area forest type remote sensing classification. The application first fuses the horizontal spatial and spectral features (spectrum, phenology, texture, etc.) and the vertical structure features (canopy height, growth rate) of the forest canopy, solves the limitation of traditional methods in large-scale forest type remote sensing classification which only rely on single-dimensional features, extracts the vertical structure features of the forest canopy by fusing the spaceborne laser radar (GEDI / ICESat-2) and optical remote sensing (Landsat / Sentinel-2) data, replaces the traditional method which relies on airborne LiDAR, and effectively reduces the cost of large-area forest monitoring. The “growth rate feature” (annual average height / spectral / texture feature change rate) is proposed to represent the differences in canopy development of different forest types from the time dimension, and provides basic data support for the optimal allocation of forest resources.

[0041] Of course, implementing any product or method of the application does not necessarily require achieving all the advantages described above at the same time. BRIEF DESCRIPTION OF DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0043] Figure 1 is a flowchart of the forest canopy vertical structure feature inversion method described in the embodiment of the application;

[0044] Figure 2 is a flowchart of the large-area forest classification method described in the embodiment of the application;

[0045] Figure 3 is a general accuracy chart of forest type remote sensing classification based on different classification feature combinations in the application example;

[0046] Figure 4 is a classification accuracy F1 score statistical chart of different forest types based on different classification feature combinations in the application example;

[0047] Figure 5is a different forest type classification accuracy PA score statistical chart based on different classification feature combinations in the application example of the present application;

[0048] Figure 6 is a different forest type classification accuracy UA score statistical chart based on different classification feature combinations in the application example of the present application;

[0049] Figure 7 is a forest type remote sensing classification detail comparison chart based on different classification feature combinations in the application example of the present application. DETAILED DESCRIPTION

[0050] The present application inventors discovered the above problems and conducted a detailed study on the existing large-area forest type remote sensing classification methods. The study found that the forest type remote sensing classification method based on satellite remote sensing data can be roughly divided into four categories:

[0051] The first method is forest type remote sensing classification based on the difference in spectral response characteristics (spectral features) of the forest canopy; that is, forest type classification is performed according to the difference in spectral reflectance characteristics of the forest canopy with respect to electromagnetic wavelength. For example, some scholars use the spectral feature differences of forest types on QuickBird images to classify coniferous forests, broadleaf forests, and coniferous-broadleaf mixed forests; or use the subtle spectral feature differences of different forest types on hyperspectral images (Earth Observing-1, EO-1) to perform fine classification of forest types in the western Himalayas. This method relies only on the spectral feature differences of forest types in single temporal remote sensing images for forest type remote sensing classification, and has very high requirements for image quality, spatial resolution, and spectral resolution. However, due to the limitations of complex stand structure and sensor spectral resolution, there is a serious spectral confusion phenomenon between forest types, and single temporal optical remote sensing data often cannot achieve fine classification of forest types.

[0052] The second method is forest type remote sensing classification based on phenological features, that is, multi-temporal remote sensing data is used to extract the seasonal changes (phenology) formed by forest vegetation in the long-term adaptation to environmental changes, and the forest types are classified according to the differences in these phenological features. For example, some scholars use spring, summer, and autumn Sentinel-2 data to extract forest type phenological features, and classify forest types according to the differences in phenological features; or combine autumn and spring Sentinel-2 data to extract forest phenological features, and classify forest types according to the differences in phenological features. Compared with the forest type classification method relying only on single temporal spectral features, the phenological features effectively represent the annual growth rhythm of different forest vegetation formed in the long-term adaptation to environmental changes, providing an important research perspective for large-scale forest type remote sensing classification. However, this method ignores the influence of spectral and spatial texture features on forest type remote sensing classification.

[0053] The third method is forest type remote sensing classification based on the difference of forest canopy spatio-temporal spectral features, that is, using hyperspectral remote sensing data to enhance the spectral resolution of multispectral remote sensing data, thereby obtaining optical remote sensing data with high spatial, temporal and spectral resolution, to extract fine forest canopy phenology, spatial texture and spectral features, and then performing forest type classification based on the difference of these features. For example, some scholars use airborne hyperspectral imagery to enhance the spectral resolution of Sentinel-2 imagery to obtain remote sensing imagery with high temporal and spectral resolution, and then perform forest type classification based on the difference of spectral, phenological and spatial texture features of forest canopy; or use the high spectral resolution of CHRIS-Proba data to enhance the spectral resolution of Landsat8 OLI data to obtain remote sensing imagery with high spatial and spectral resolution, and then use the spectral, phenological and spatial texture features and the spectral angle mapper (SAM) to perform forest type classification. Compared with the forest type remote sensing classification method using only phenological features, the fusion of spatio-temporal spectral features reduces the influence of spectral confusion on forest type remote sensing classification. However, this method ignores the potential influence of forest canopy vertical structure features on forest type remote sensing classification accuracy.

[0054] The fourth method is forest type remote sensing classification by fusing forest canopy horizontal spatio-temporal spectral features and vertical structure features; that is, first extracting forest canopy spatio-temporal spectral features (spectrum, spatial texture and phenology) using optical remote sensing imagery, then extracting forest canopy vertical structure features such as height, shape and density using LiDAR remote sensing data, and finally fusing the forest canopy horizontal spatio-temporal spectral features and vertical structure features to perform forest type remote sensing classification. For example, some scholars first extract forest canopy vertical structure features such as height, shape and echo intensity using airborne LiDAR data, then extract forest canopy spectral features using satellite multispectral data, and finally perform tree species classification based on the difference of forest canopy spectral and vertical structure features; or first extract forest type phenological, spatial texture and spectral features using multi-temporal (spring, summer and autumn) satellite optical remote sensing imagery (GeoEye, Pleiades and WorldView2), then accurately extract forest canopy three-dimensional spatial structure features using airborne LiDAR data in the leafing period, and finally perform forest type remote sensing classification based on the difference of forest canopy phenological, spectral, texture and three-dimensional spatial structure features. Compared with the forest type remote sensing classification method using only forest canopy horizontal spatio-temporal spectral features, the addition of forest canopy vertical structure features significantly improves the accuracy of regional forest type remote sensing classification. However, this method relies on high-density airborne LiDAR data to extract accurate forest canopy vertical structure features, and is limited by the cost of data acquisition, making it difficult to meet the needs of large-scale forest type remote sensing classification.

[0055] Through the above analysis, it is found that the fusion of forest canopy spatiotemporal spectral characteristics (phenology, spatial texture, and spectrum) significantly improves the accuracy of forest type classification based on satellite remote sensing images. However, the spatiotemporal spectral characteristics currently used only represent the differences in forest canopy level spatiotemporal spectral characteristics, ignoring the influence of forest vertical structure characteristics on forest type remote sensing classification. In fact, different forest vegetation types have significant differences in vertical structure parameters (such as tree height, crown shape, growth rate, etc.) due to their comprehensive adaptation to the ecological environment during the long process of phylogenetic development, especially the competitive pressure from surrounding trees, which can complement the feature classification ability of canopy level spatiotemporal spectral characteristics to some extent, theoretically further improving the separability between forest types and thus improving the classification accuracy. Although existing research has confirmed that the fusion of canopy three-dimensional structure features extracted by airborne LiDAR and horizontal spatiotemporal spectral features can achieve high forest classification accuracy within a small scale, due to the high cost of airborne LiDAR data acquisition, this method is difficult to apply to large-scale forest classification. Therefore, exploring a method for extracting forest canopy vertical structure features based on satellite remote sensing data and fusing canopy horizontal spatiotemporal spectral and vertical structure features for large-scale forest type remote sensing classification is still a scientific problem that needs to be solved.

[0056] It should be noted that the defects in the above prior art solutions are the result of the inventors' practice and careful study, therefore, the discovery process of the above problems and the solutions proposed by the embodiments of the present application to solve the above problems should be the contribution of the inventors to the present application during the process of the present application.

[0057] The technical solutions in the embodiments of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0058] It should be noted that: similar reference numerals and letters represent similar items in the following drawings, therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. In the description of the present application, the terms "first", "second", "third", "fourth", etc. are only used for differentiation, and cannot be understood as indicating or implying relative importance.

[0059] Based on the above in-depth analysis, this invention provides a method for inverting the vertical structure features of the forest canopy, a method for classifying large-area forests, and a system. Different forest types, through long-term phylogenetic development, have undergone comprehensive adaptation to the ecological environment, especially to competitive pressures from surrounding forests, resulting in significant differences in canopy height and growth rate characteristics. For example, oak forests generally grow slower than birch forests; while the growth rate of *Pinus yunnanensis* is generally higher than that of *Pinus yunnanensis* and *Pinus armandii*, and the canopy height of *Abies spp.* is significantly higher than that of *Pinus yunnanensis* and *Pinus armandii*. Forest canopy height describes the differences in the three-dimensional spatial structure of the forest canopy from the perspective of canopy spatial structure, while growth rate characteristics describe the differences in the growth and development characteristics of forest types from the perspective of time. The fusion of these two spatiotemporal structural features of the forest canopy can, to some extent, complement the classification capabilities of spectral, spatial texture, and phenological features extracted from optical remote sensing data, thereby helping to improve the classification accuracy of plateau and mountain forest types. To address the lack of characterization of large-scale forest vertical structure parameters in existing research, this invention integrates spaceborne optical and lidar data to extract canopy height and growth rate features to comprehensively characterize canopy vertical structure features, and integrates canopy horizontal features to form comprehensive spatiotemporal features for remote sensing classification of large-area forest types in complex environments.

[0060] like Figure 1 As shown, the canopy vertical structure feature inversion method provided in this embodiment of the invention includes the following steps:

[0061] Step S101: Determine the monitoring area, collect sample point data and long-term optical remote sensing data from two different years within a long time series, and infer the spatial distribution map of forest canopy height for the two years respectively.

[0062] Step S102: Based on the spatial distribution map of canopy height, extract the pixels in the second year that show an increase in forest canopy height compared to the first year as a height sample set, and use the pixels that do not show an increase as a feature sample set.

[0063] Step S103: Calculate the height difference and feature difference between the second year and the first year based on the height sample set and feature sample set respectively, construct a relationship model between height difference and feature difference, and obtain forest canopy height data over a long time series.

[0064] Step S104, obtain remote sensing images in the monitoring area in the long time series in Google Earth Engine (GEE), and after preprocessing, represent the spectral characteristics of the forest canopy by Normalized Difference Vegetation Index (NDVI), and represent the spatial texture characteristics of the forest canopy by Sum of Squares: Variance (svar) calculated by Gray-Level Co-occurrence Matrix (GLCM) of Spectral variability vegetation index (SVVI).

[0065] In this step, the SVVI_svar is calculated by the SVVI index and the GLCM method.

[0066] Step S105, fill in the forest canopy height data, NDVI and SVVI_svar corresponding to the missing years of the pixels by using an interpolation method, to obtain a complete forest canopy height data set, an NDVI set and an SVVI_svar set.

[0067] Step S106, obtain forest disturbance detection data in the long time series in the monitoring area, and take the year when each pixel has forest disturbance as the growth starting point of the pixel.

[0068] Step S107, based on the forest canopy height data set, the NDVI set and the SVVI_svar set, obtain the canopy height, spectral reflectance and spatial texture of each pixel from the growth starting point to the last year in the long time series, and quantitatively describe the growth trajectory of the forest type.

[0069] Step S108, based on the forest type growth trajectory of each pixel, extract the annual growth speed characteristics of the forest canopy spectrum, spatial texture and canopy height from the year when each pixel has forest disturbance to the last year in the long time series or when the spectral reflectance of the forest canopy reaches saturation, and use the characteristics to represent the forest canopy growth speed characteristics of different forest types.

[0070] In this step, the growth speed characteristics of the forest canopy are represented by inverting the canopy spectrum, spatial texture and canopy height. The growth and development of the forest not only shows the increase of the tree height, but also shows the occupation and distribution heterogeneity of the forest canopy elements such as leaves and branches in the three-dimensional space, i.e. the changes of the forest canopy spectrum and spatial texture characteristics.

[0071] Step S109, represent the vertical structure characteristics of the forest canopy by the canopy height characteristics and the growth speed characteristics.

[0072] Based on the forest canopy vertical structure characteristics obtained by the above forest canopy vertical structure characteristic inversion method, the embodiment of the present application further provides a large-area forest classification method. Figure 2 As shown in the figure, the method comprises the following steps:

[0073] Step S1, determine the monitoring area and the long time interval, and obtain the multi-source remote sensing data of the monitoring area, wherein the multi-source remote sensing data comprises SRTM DEM data, Landsat series data in the long time interval, spaceborne LiDAR data (GEDI and ICESat-2 ATLAS), Sentinel-1 / 2 data, meteorological data, soil data and forest type field survey data.

[0074] In this step, the multi-source remote sensing data is basically obtained from satellite remote sensing data; the long time interval is a time period, such as 1986-2023.

[0075] Step S2, extracting forest canopy horizontal spatiotemporal spectral features by using multi-source remote sensing data.

[0076] In this step, the extracted forest canopy horizontal spatiotemporal spectral features include spectrum, phenology, texture, polarization, geographical environment, etc. Each kind of horizontal spatiotemporal spectral feature contains several different specific features or indexes, and the specific features under each kind of feature are listed in Table 1, and the vegetation indices used in Table 1 and the corresponding calculation formulas are listed in Table 2.

[0077] Table 1 Extraction of forest canopy horizontal spatiotemporal spectral features

[0078]

[0079] Table 2 Vegetation indices used in the present application and corresponding calculation formulas

[0080]

[0081] In Table 2, represent green, blue, red, near-infrared, short-wave infrared 1, short-wave infrared 2, red edge band 1, red edge band 2, red edge band 3 and red edge band 4 respectively; represents the standard deviation of the blue, green, red, near-infrared, SWIR1 and SWIR2 bands; represents the standard deviation of the NIR, SWIR1 and SWIR2 bands. VV and VH represent the backscattering coefficients of vertical polarization (single polarization, both transmission and reception are vertical) and horizontal polarization (dual polarization, transmission direction is opposite to vertical, and reception is horizontal direction) respectively.

[0082] In this step, the texture features are extracted by using RedEdge1 band and 18 texture features calculated by Gray level co-occurrence matrix (GLCM) to comprehensively represent the spatial texture feature difference of forest vegetation canopy.

[0083] In the extraction of the phenology features, 14 statistical values of the intra-annual time series data are selected, including 10-day interval Red-edge Position (REP), comprehensive statistical features of time series data, intra-annual Vertical-Vertical (VV), Vertical-Horizontal (VH), standard deviation of Modified radar vegetation index (mRVI), maximum difference of intra-annual and winter (from December to March) Normalized Difference Vegetation Index (NDVI) (NDVI_maxsummer) and mRVI maximum difference composite image of summer and winter (mRVI_summerwinter) to jointly represent the phenology feature difference between different forest types; wherein, for the extraction of the comprehensive statistical features of REP time series data, linear interpolation method and Savitzky-Golay filtering algorithm are used to generate 10-day interval REP time series data set within a year; and the mean, standard deviation, maximum, minimum, median, range of maximum and minimum (RMM), coefficient of variation (cv), first quartile (Q25), third quartile (Q75), the interquartile range of Q75 and Q25 (IRQ3Q1), the interquartile range of Q25 and minimum (IRQ25Qmin), the interquartile range of maximum and Q75 (IRQmaxQ75) are calculated by using all REP time series data to represent the phenology feature difference of different forest types.

[0084] When extracting the geographical environment features, three terrain features of elevation, slope and aspect are extracted based on SRTM DEM data; four climate features of annual mean temperature, seasonal mean temperature, annual mean precipitation and seasonal mean precipitation are extracted based on global 1km resolution, monthly scale climate products (WorldClim_version 2.1) in a long time interval; a soil moisture response index (SMRI) is calculated by using a water loss rate model based on the soil moisture data product (Soil Moisture of China by the in situ data, version 1.0, SMCI1.0) of China 1km resolution, daily scale and 10-100cm depth in a year, which is used to quantify the ability of soil to provide water required for the growth and development of forest types; and all geographical environment covariates are resampled to a spatial resolution of 30m by using a bilinear interpolation method.

[0085] The calculation formula of the seasonal mean temperature (precipitation) is as follows:

[0086] (1)

[0087] In formula (1), MS represents the seasonal mean temperature (precipitation), and respectively represent the standard deviation and mean value of the monthly scale temperature (precipitation).

[0088] Step S3: acquiring the vertical structure features between forest types by using the forest canopy vertical structure feature inversion method as described above.

[0089] Step S4: fusing the forest canopy horizontal spatio-temporal spectral features and the vertical structure features, and performing large-scale forest type remote sensing classification by using the forest type samples investigated in the field and a machine learning method.

[0090] Further, the step specifically includes:

[0091] Step S41: fusing the extracted forest canopy horizontal spatio-temporal spectral features and the vertical structure features to jointly constitute a feature space for forest type remote sensing classification.

[0092] Step S42: based on the forest canopy horizontal spatio-temporal spectral features and the vertical structure features, calculating feature correlation and feature importance by using an association hierarchical clustering method and a random forest algorithm (RandomForest, RF) to select classification features beneficial to forest type remote sensing classification in the monitoring region, and collecting all the beneficial classification features as an optimal classification feature combination.

[0093] In this step, the beneficial classification features are taken as the coniferous forest and the broad-leaved forest, for example, when the forest types include the coniferous forest and the broad-leaved forest, the beneficial classification features of the forest canopy horizontal structure corresponding to the two types are shown in Table 3 and Table 4.

[0094] Table 3 Preferred classification features of the coniferous forest area

[0095]

[0096] Table 4 Preferred classification features of the broad-leaved forest area

[0097]

[0098] Step S43, based on the optimal combination of classification features, the forest type remote sensing classification is performed by using the RF classifier.

[0099] Step S5, the forest type remote sensing classification accuracy is quantitatively evaluated by using the forestry field investigation sample data of the target year.

[0100] In this step, based on the verification sample of the field investigation, five indexes including the commonly used overall accuracy (Overall Accuracy, OA), Kappa coefficient, producer accuracy (Producer’s Accuracy, PA), user accuracy (User’s Accuracy, UA) and F1 score (F1 score) are selected to quantitatively evaluate the effectiveness of the forest canopy vertical structure features extracted by the present application on the remote sensing classification of the plateau mountain forest type. The calculation formulas of these indexes are as follows:

[0101] (2)

[0102] (3)

[0103] (4)

[0104] (5)

[0105] (6)

[0106] In formula (2)-(6), is the total number of forest types, is the total number of verification samples, represents the elements of the diagonal line of the confusion matrix, that is, the total number of samples correctly classified in each forest type classification, and The column total and the row total of each forest type represent the total number of samples of misclassification and missed classification, respectively; the larger the values of OA and Kappa, the higher the overall classification accuracy of the regional forest types; the larger the values of PA, UA and F1score, the higher the classification accuracy of each forest type.

[0107] Based on the same idea, the embodiment of the present application also provides a large-area forest classification system, which comprises a data acquisition module, a horizontal spatio-temporal spectral feature extraction module, a vertical structure feature extraction module, a remote sensing classification module and an evaluation module.

[0108] The data acquisition module is configured to determine a monitoring area and a long-time interval, and acquire multi-source remote sensing data of the monitoring area, wherein the multi-source remote sensing data comprises SRTM DEM data, Landsat series data in the long-time interval, spaceborne LiDAR data, Sentinel-1 / 2 data, meteorological data, soil data and field survey data of forest types.

[0109] The horizontal spatio-temporal spectral feature extraction module is configured to extract forest canopy horizontal spatio-temporal spectral features by using the multi-source remote sensing data.

[0110] The vertical structure feature extraction module is configured to obtain vertical structure features between forest types by using the forest canopy vertical structure feature inversion method as described above.

[0111] The remote sensing classification module is configured to fuse the forest canopy horizontal spatio-temporal spectral features and the vertical structure features, and perform large-scale forest type remote sensing classification by using forest type samples obtained through field survey and a machine learning method.

[0112] The evaluation module is configured to quantitatively evaluate the forest type remote sensing classification accuracy by using forest field survey sample data of a target year.

[0113] The modules in the embodiment are implemented by a processor, and a memory is appropriately added when storage is needed. The processor can be, but is not limited to, a microprocessor (MPU), a central processing unit (CPU), a network processor (NP), a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The memory can include a random access memory (RAM) and can also include a non-volatile memory (NVM), such as at least one disk memory. Optionally, the memory can also be at least one storage device located away from the aforementioned processor.

[0114] In the above embodiments, all or part of the embodiments can be realized by software, hardware, firmware or any combination thereof. When realized by software, all or part of the embodiments can be realized in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network or other programmable device. The computer instructions can be stored in a computer readable storage medium or transferred from one computer readable storage medium to another computer readable storage medium, for example, the computer instructions can be transferred from one website, computer, server or data center to another website, computer, server or data center through wired (for example, coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (for example, infrared, wireless, microwave, etc.).

[0115] In addition, it should be noted that the large-area forest classification system and the large-area forest classification method described in the embodiments are corresponding, and the description and limitation of the method are also applicable to the system, which will not be repeated here.

[0116] The large-area forest classification method and system described in the embodiments of the present application are applied to the forest type classification of a monitoring area in Yunnan Province, training samples and verification samples are generated based on the monitoring area, and the forest type remote sensing classification accuracy and results of different classification feature combinations are compared and analyzed.

[0117] For classification accuracy, the preferred features shown in Tables 3 and 4 are selected based on the training samples, the forest type classification is performed using the random forest classifier, and the effectiveness of different classification feature combinations on forest type remote sensing classification is quantitatively evaluated using the verification samples, and the evaluation results are as follows: Figure 3The overall fusion of the time-space-spectrum-polarization characteristics of the forest canopy significantly improved the classification accuracy of the forest types in the plateau mountainous area, but the classification contribution of different characteristics was different. Due to the similar waveform of the spectral reflectance curve of different forest types, but they showed certain differences in the backscattering coefficient, the OA and Kappa coefficients of the forest type remote sensing classification of Yunnan Province by fusing Sentinel-1 and Sentinel-2 data were 65.38% and 0.5770, respectively. The texture characteristics had little contribution to the remote sensing classification accuracy of forest types. Compared with using only spectral and polarization characteristics, the addition of texture characteristics increased the OA and Kappa by 0.14% and 0.0005, respectively. The comprehensive adaptation of forest types to the growth environment caused great differences in the growth rhythm characteristics of different forest types. In the process of forest type remote sensing classification, the addition of phenology characteristics can significantly improve the tree species classification accuracy; compared with using only spectral, polarization and texture characteristics, the addition of phenology characteristics increased the OA and Kappa by 1.13% and 0.0227, respectively. Forest vegetation always exchanges matter and energy with the surrounding environment during growth and development, and topography, climate, hydrology and soil and other geographical environmental factors significantly affect the spatial distribution pattern of forest types. Therefore, different forest types have significant differences in geographical environmental factors, such as rubber mainly distributed in low-altitude areas with sufficient water and heat conditions, and spruce mainly distributed in high-altitude areas. The addition of geographical environmental characteristics significantly improved the remote sensing classification accuracy of forest types in the plateau mountainous area, and the OA and Kappa increased by 12.55% and 0.1442, respectively, compared with not adding geographical environmental characteristics. The temporal and spatial structure characteristics of forest canopy can effectively improve the remote sensing classification accuracy of forest types in the plateau mountainous area, and compared with using only spectral, polarization, texture, phenology and geographical environment characteristics of forest canopy, the fusion of time-space-spectrum-polarization characteristics of forest canopy significantly improved the remote sensing classification accuracy of forest types, and the OA and Kappa coefficients increased by 3.29% and 0.0393, respectively.

[0118] In order to verify the contribution of different classification characteristic combination methods to different forest type remote sensing classification, this embodiment compared and analyzed the change trend of classification accuracy of different forest types under different classification characteristic combination methods. For example Figures 4-6As shown, the fusion of forest canopy time-space-spectrum-polarization features significantly improves the classification accuracy of various forest types as a whole. However, due to the differences in tree species composition and the complexity of the plant community of different forest types, the classification accuracy of each forest type under different feature combination methods varies greatly. Overall, the classification accuracy of forest types is negatively correlated with the complexity of tree species composition and community structure; for example, the classification accuracy of coniferous forest types is significantly higher than that of broad-leaved forest types, and for example, the classification accuracy of Yunnan pine and Simao pine pure forests is significantly higher than that of other coniferous forests, and for example, rubber, as a typical economic forest in Yunnan Province, has a single tree species composition, and its classification accuracy is significantly higher than that of oak, birch, and other broad-leaved forests with complex community composition. However, the contribution of different classification feature combinations to the classification performance of different tree species varies.

[0119] First, in the classification of coniferous forest types, the classification accuracy of Yunnan pine generally shows a trend of first decreasing and then increasing with the increase in the number of classification features; among them, the mapping accuracy of Yunnan pine with the fusion of spectral, polarization and texture features is the lowest, with an F1 score of 0.7534. Compared with only using spectral and polarization features, the addition of texture features causes a certain degree of feature redundancy, and the classification accuracy of Yunnan pine decreases slightly, with a decrease in F1 score and UA of 0.0024 and 1.87%, but an increase in PA of 1.54%. This proves that the addition of texture features slightly increases the misclassification rate of Yunnan pine, but significantly reduces the omission rate. The fusion of forest canopy time-space-spectrum-polarization features significantly improves the classification accuracy of Yunnan pine, with an increase in F1 score and PA of 0.0121 and 2.3%, respectively. The classification accuracy of Simao pine generally shows a trend of first increasing and then decreasing and then increasing with the increase in classification features, among them, the classification accuracy after adding forest canopy time-space structure features is the highest, with F1 score, UA and PA being 0.9596, 94.41% and 97.57%, respectively. Compared with before adding forest canopy time-space structure features, the F1 score and UA are increased by 0.0081 and 2.06%, respectively. The classification accuracy of Yunnan spruce increases continuously with the increase in classification features, and the addition of forest canopy time-space structure features significantly improves the classification accuracy of Yunnan spruce. Compared with before adding forest canopy time-space structure features, the F1 score, UA and PA of Yunnan spruce with the fusion of forest canopy time-space-spectrum-polarization features are increased by 0.0377, 3.57% and 2.81%, respectively. The classification accuracy of other coniferous forest types increases continuously with the increase in classification features, and the F1 score, UA and PA of other coniferous forests are increased by 0.0262, 3.6% and 1.54%, respectively, after the fusion of forest canopy time-space-spectrum-polarization features.

[0120] Secondly, in the remote sensing classification of broad-leaved forest forest types, oak as a typical hard broad forest type grows slowly and contains many forest types. The classification accuracy of plateau mountain oak generally increases first and then decreases with the increase of classification accuracy. Among them, the classification accuracy of oak fused with forest canopy time-space-spectrum-polarization characteristics is the highest, and the F1 score, UA and PA are 0.7232, 84.55% and 63.18% respectively. Compared with the feature of not adding forest canopy time-space structure, the F1 score, UA and PA are increased by 0.0618, 10.79% and 3.23% respectively. As a typical soft broadleaf tree species (group) in Yunnan Province, birch grows faster, and the addition of forest canopy time-space structure characteristics significantly improves the classification accuracy of birch species. Compared with the feature of not adding forest canopy time-space structure, the F1 score, UA and PA of the feature of fusing forest canopy time-space-spectrum-polarization characteristics are increased by 0.1209, 9.86% and 15.34% respectively. The classification accuracy of rubber decreases first and then increases with the increase of classification features. Among them, the classification accuracy of rubber fused with forest canopy time-space-spectrum-polarization characteristics is the highest, and the F1 score, UA and PA are 0.9011, 82.00% and 100% respectively. Compared with the feature of not adding forest canopy time-space structure, the PA is increased by 4.65%. The classification accuracy of other broad-leaved forests increases first and then decreases and then increases with the increase of classification accuracy. Among them, the classification accuracy of rubber fused with forest canopy time-space-spectrum-polarization characteristics is the highest, and the F1 score, UA and PA are 0.3648, 30.37% and 45.67% respectively. Compared with the feature of not adding forest canopy time-space structure, the PA is increased by 7.00%.

[0121] In addition, based on the remote sensing classification results of plateau mountain forest types of different feature combinations, the present embodiment also carries out comparative analysis. The fusion of forest canopy time-space-spectrum-polarization characteristics significantly improves the classification accuracy of plateau mountain forest types.

[0122] In addition, the present application further compares the effectiveness of the fusion of forest canopy time-space-spectrum-polarization characteristics on the remote sensing classification of plateau mountain forest types by using the photos taken on the spot, as shown in Figure 7 It can be seen from Figure 7 that the remote sensing classification mapping results of forest types in Yunnan Province based on different classification feature combinations have great differences in detail scale. From Figure 7It can be seen that the classification of Pinus kesiya var. lasiocarpa is seriously confused with Pinus yunnanensis. After fusing the geographical environment features and the spatio-temporal structure features of forest canopy, the misclassification error of Pinus kesiya var. lasiocarpa is significantly improved. Compared with the use of spectral, polarization, texture and phenology features alone, the misclassification error of Pinus kesiya var. lasiocarpa is reduced by 47.94% after fusing these features with geographical environment features. Compared with the addition of spatio-temporal structure features of forest canopy, the UA of Pinus kesiya var. lasiocarpa is increased by 2.06% after fusing the spatio-temporal-spectral-polarization features of forest canopy, and the misclassification error is significantly reduced. The extraction of rubber plantation based on spectral, polarization, texture and phenology features presents the misclassification of oak and other broad-leaved forests. Since rubber plantations are mainly distributed in low-altitude areas with sufficient water and heat, the misclassification rate of rubber plantations is significantly reduced after adding geographical environment features. Compared with the fusion of geographical environment features, the UA is increased by 46%. The spectral, polarization, texture, phenology and geographical environment features of forest canopy of oak are similar to those of birch and other broad-leaved forests. However, due to the slow growth of oak, there are significant differences in the spatio-temporal structure features of forest canopy between them.

[0123] Compared with the addition of spatio-temporal structure features of forest canopy, the UA is increased by 3.50% after fusing the spatio-temporal-spectral-polarization features of forest canopy. Figure 7 It can be seen that the other coniferous forests in high-altitude areas are misclassified as Pinus yunnanensis and Abies fabri. After adding the spatio-temporal structure features of forest canopy, the misclassification rate and the omission rate of other coniferous forests are significantly reduced, and the UA and PA are increased by 3.60% and 1.54%, respectively. Abies fabri is largely misclassified as Pinus yunnanensis and other coniferous forests. Considering the unique growth environment of Abies fabri, the classification accuracy of Abies fabri is significantly improved by fusing spectral, polarization, texture, phenology and geographical environment features. Compared with the addition of geographical environment features, the UA and PA are increased by 22.29% and 13.10%, respectively. In addition, Abies fabri grows slowly and its forest canopy height is significantly higher than that of other coniferous forests. Therefore, compared with the addition of spatio-temporal structure features of forest canopy, the UA and PA of Abies fabri are increased by 17.62% and 2.81%, respectively, after fusing the horizontal spatio-temporal-spectral and vertical structure features. The classification mapping of birch forest before adding the spatio-temporal structure features of forest canopy presents a large number of confusions with oak and other broad-leaved forests. Since birch forest belongs to typical soft broad-leaved forest and grows fast, the classification accuracy of birch is significantly improved after fusing the spatio-temporal-spectral-polarization features of forest canopy, and the UA and PA are increased by 9.66% and 15.34%, respectively. Therefore, the classification accuracy of each forest type in the plateau mountainous area is significantly improved by fusing the spatio-temporal-spectral-polarization features of forest canopy, and the mapping accuracy of birch, oak and Abies fabri is most significantly improved.

[0124] From the above, the forest canopy vertical structure feature inversion method, the large-area forest classification method and the system provided by the embodiment of the present application improve the classification effect of the large-area forest type, and improve the accuracy and precision of the classification.

[0125] The above description is merely the preferred embodiments of the present application and the description of the applied technical principles, and is not intended to limit the scope of the claimed present application, but merely represents the preferred embodiments of the present application. Those skilled in the art should understand that the scope of the present application involves the technical solutions formed by the specific combinations of the technical features described above, and should also cover other technical solutions formed by any combination of the technical features described above or their equivalent features without departing from the inventive concept. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.

Claims

1. A method for retrieving vertical structure features of forest canopy, characterized in that, The method comprises the following steps: Step S101, determining a monitoring area, collecting sample point data and long-time optical remote sensing data of two different years in a long time sequence, and respectively inferring spatial distribution maps of forest canopy height of the two years; Step S102, extracting, according to the spatial distribution maps of the forest canopy height, pixels with increased forest canopy height in the second year than in the first year as a height sample set, and pixels without increase as a feature sample set; Step S103, respectively calculating height difference and feature difference corresponding to the second year and the first year, constructing a relationship model of the height difference and the feature difference, and obtaining forest canopy height data in the long time sequence; Step S104, obtaining remote sensing images in the long time sequence of the monitoring area, and taking normalized difference vegetation index (NDVI) to represent spectral characteristics of the forest canopy after preprocessing, and taking spectral variation vegetation index variance (SVVI_svar) to represent spatial texture characteristics; Step S105, filling in missing values to obtain a complete forest canopy height data set, an NDVI set and an SVVI_svar set; Step S106, obtaining forest disturbance detection data in the long time sequence of the monitoring area, and taking a year in which forest disturbance occurs as a growth development starting point; Step S107, based on the forest canopy height data set, the NDVI set and the SVVI_svar set, obtaining canopy height, spectral reflectance and spatial texture of each pixel from the growth development starting point to the last year in the long time sequence, and quantitatively describing a growth trajectory of a forest type; Step S108, respectively extracting, from the year in which forest disturbance occurs to the last year in the long time sequence or to a time when spectral reflectance of the forest canopy reaches saturation, spectral characteristics, spatial texture characteristics and canopy height characteristics of the forest canopy to represent forest canopy growth speed characteristics; Step S109, representing forest canopy vertical structure characteristics by using the canopy height characteristics and the growth speed characteristics together.

2. The method according to claim 1, wherein, In step S103, height difference and feature difference corresponding to the second year and the first year are respectively calculated according to the height sample set and the feature sample set.

3. A method of large area forest classification, characterized by, The method comprises: Step S1, determining a monitoring area and a long time sequence interval, and obtaining multi-source remote sensing data of the monitoring area; Step S2, extracting canopy horizontal spatio-temporal spectral characteristics by using the multi-source remote sensing data; Step S3, obtaining vertical structure characteristics between forest types by using the canopy vertical structure characteristic inversion method according to any one of claims 1-2; Step S4, fusing the canopy horizontal spatio-temporal spectral characteristics and the vertical structure characteristics, and performing large-scale forest type remote sensing classification by using forest type samples investigated in the field and a machine learning method; Step S5, quantitatively evaluating mapping accuracy by using forest field investigation sample data of a target year.

4. The method of claim 3, wherein, In step S1, the multi-source remote sensing data comprises SRTM DEM data, Landsat series data in the long time sequence interval, spaceborne LiDAR data, Sentinel-1 / 2 data, meteorological data, soil data and forest type field investigation data.

5. The method of claim 3, wherein, Step S4 specifically comprises: Step S41, fusing the extracted canopy horizontal spatio-temporal spectral characteristics, the canopy height characteristics and the growth speed characteristics to jointly constitute a feature space for forest type remote sensing classification; Step S42: Based on the forest canopy level spatio-temporal spectral features and vertical structure features, the correlation and importance of features are calculated by using the correlation hierarchical clustering method and the random forest algorithm RF to select the classification features beneficial to the remote sensing classification of forest types in the monitoring area, and all the beneficial classification features are combined as the optimal classification feature combination; Step S43, based on the optimal classification feature combination, the RF classifier is used for remote sensing classification of forest types.

6. The large area forest classification method according to claim 3, wherein, The forest canopy level spatio-temporal spectral features extracted in step S3 include spectrum, phenology, texture, polarization and geographic environment.

7. The method of claim 6, wherein, When extracting the texture features, 18 texture features calculated by RedEdge1 band and gray level co-occurrence matrix are used to comprehensively represent the spatial texture feature differences of forest vegetation canopy.

8. The large area forest classification method according to claim 6, characterized in that, When extracting the phenology features, 14 statistical values of the year-round time series data are selected; the statistical values include: 10-day interval red edge position index REP, comprehensive statistical features of time series data, year-round vertical polarization VV, vertical-horizontal polarization VH, standard deviation of modified vegetation index mRVI, maximum difference of year-round and winter normalized vegetation index NDVI_maxsummer, and mRVI maximum value difference of summer and winter mRVI_summerwinter; wherein, for the extraction of the comprehensive statistical features of the time series data, the linear interpolation method and the Savitzky-Golay filtering algorithm are used to generate the REP time series data set with 10-day interval in a year, and 12 features including mean, standard deviation, maximum value, minimum value, median, range, coefficient of variation, 0.25 quantile, 0.75 quantile, interquartile range 1, interquartile range 2, and interquartile range 3 are calculated based on all REP time series data to represent the phenology feature differences of different forest types.

9. The large area forest classification method according to claim 6, characterized in that, When extracting the geographic environment features, three terrain features of elevation, slope and aspect are extracted; four climate features of annual mean temperature, seasonal mean temperature, annual mean precipitation and seasonal mean precipitation are extracted based on global 1km resolution, monthly scale climate products in a long time interval; based on the 1km resolution, daily scale, 10-100cm depth soil moisture data product in a year, the soil moisture response index SMRI is calculated by using the water loss rate model to quantify the ability of soil to provide water required for the growth and development of forest types; and the bilinear interpolation method is used to resample all geographic environment covariates to 30m spatial resolution.

10. A large area forest classification system characterized by, The system comprises: a data acquisition module, a horizontal spatio-temporal spectral feature extraction module, a vertical structure feature extraction module, a remote sensing classification module and an evaluation module; wherein, The data acquisition module is used to determine the monitoring area and the long time series interval, and to acquire multi-source remote sensing data of the monitoring area; The horizontal spatio-temporal spectral feature extraction module is used to extract the forest canopy level spatio-temporal spectral features by using the multi-source remote sensing data; The vertical structure feature extraction module is used to acquire the vertical structure features between forest types by using the forest canopy vertical structure feature inversion method according to any one of 1-2. The remote sensing classification module is used to perform remote sensing classification of forest types by using the optimal classification feature combination obtained by the vertical structure feature extraction module. The remote sensing classification module is used for fusing forest canopy horizontal space-time spectral characteristics and vertical structure characteristics, and using field investigation forest type samples and a machine learning method to perform large-scale forest type remote sensing classification. The evaluation module is used for quantitatively evaluating the mapping precision by using the forestry field investigation sample data of the target year.

Images