A quantitative analysis method for surface habitat factors

Through high-resolution remote sensing image and semi-supervised recursive classification technology, the problem of insufficient analytical accuracy and efficiency of habitat factor data is solved, and efficient and multi-dimensional habitat factor data acquisition and analysis is achieved, supporting ecological environment research at the medium-microscale scale.

CN113361350BActive Publication Date: 2024-11-12SOUTHEAST UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202110573953.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-05-25
Publication Date
2024-11-12
Estimated Expiration
2041-05-25

AI Technical Summary

Technical Problem

The existing technology has insufficient accuracy, efficiency and scientificity in the acquisition and analysis of habitat factor data parameters, making it difficult to process composite ecological parameter data, and the research is mostly based on macro scales, which is difficult to support ecological environment analysis at the medium-micro scale, resulting in ecological construction. Lack of high-precision guidance.

Method used

Using high-resolution remote sensing image and semi-supervised recursive classification technology, a high confidence sample library is constructed through the maximum likelihood model and principal component analysis method to realize the overall analysis and dimensionality reduction of multi-dimensional data of habitat factor objects, and improve the analytical accuracy and efficiency.

Benefits of technology

It realizes habitat factor data acquisition with high efficiency, high dimensions and high refresh rate, reduces low correlation parameter interference, improves habitat factor analysis accuracy and system operation efficiency at the medium-microscale, and supports dynamic change analysis of complex ecological environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113361350B_ABST
    Figure CN113361350B_ABST
Patent Text Reader

Abstract

The invention discloses a quantitative analysis method for surface habitat factors, which belongs to the field of digital landscape site exploration. The method improves the data collection efficiency, data dimension and analysis accuracy of habitat factor analysis by combining high spectral resolution data, high spatial resolution data and high temporal resolution data; reduces the pepper-salt phenomenon existing in previous analysis, and realizes the overall analysis of multi-dimensional data of habitat factor objects; through the principal component analysis method of habitat factor multi-dimensional data set object dimensionality reduction, the low correlation parameters of habitat factor objects are eliminated, the parameter dimension required for analysis is reduced, the analysis speed and efficiency of high-resolution remote sensing images are improved, and a low-dimensional high-confidence sample library of habitat factors is formed. The maximum likelihood method semi-supervised classification is used to perform iterative machine learning of the high-confidence sample library of habitat factors, which reduces the mutual interference of habitat factor parameters and improves the accuracy and analysis efficiency of multiple habitat factor analysis under composite parameter conditions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the field of digital landscape site exploration, and in particular relates to a method and device for analyzing surface habitat factors based on high-resolution remote sensing images and semi-supervised recursive classification technology. Background Art

[0002] Habitat factors refer to natural or artificial elements with ecological influence in the ecosystem. Their spatial forms are distributed in the form of planes, lines, or points. Their spatial distribution relationships show intertwined and mixed characteristics according to their mutual relationships and specific environments. High-precision analysis of habitat factors not only plays a key role in maintaining and protecting the local ecological pattern in the ecosystem, promoting the continuous improvement of the regional ecological pattern, and overall upward development, but also improves the scientific and ecological nature of ecological environment construction, and helps achieve high-quality landscape environment ecological civilization construction.

[0003] The current division of habitat factors in landscape environment construction has the following problems:

[0004] (1) The accuracy, efficiency and scientificity of habitat factor data parameter acquisition and analysis need to be improved, resulting in problems such as analyzing a large space with small data, analyzing a complex space with single data, and difficulty in processing comprehensive data.

[0005] The analysis of habitat factors in landscape environment construction using relevant ecological technologies and remote sensing technologies is mainly qualitative research, and the degree of quantitative analysis needs to be deepened; most of the data such as habitat factors rely on manual field surveys to collect or rely on medium and low resolution remote sensing images (resolution below 30M) for analysis. The data collection process is cumbersome, and its accuracy is difficult to improve in the spatial analysis of ecological environment at the medium and micro scales; it is difficult to coordinate and comprehensively process composite ecological parameter data, and most studies only involve some parameters, making it difficult to comprehensively and scientifically grasp the overall spatial attributes of habitat factors.

[0006] (2) Research on habitat factors is mostly based on existing models and frameworks, with the macro scale as the research object, using fixed parameter indexes for data analysis, and focusing on a single ecological sensitivity parameter index. There is little integrated and systematic research on the coordinated analysis and comprehensive analysis of multi-source complex habitat factor data. The overall analysis system's operating efficiency, accuracy, and globality need to be improved.

[0007] In summary, the current research level of habitat factors is often at the macro-scale of district, city and province, and its ability to further explore the analysis scale and analysis accuracy at the meso-micro scale needs to be strengthened; the research objects are mostly a certain habitat factor or ecological parameter; when dealing with the analysis of complex habitat factors, the parameter selection tends to focus on the selection of typical parameters for analysis, and there is a lack of consideration of its particularity when dealing with various types of meso-micro scale ecological environments; the research objects or units are mainly based on point selection based on remote sensing images, and ecological space analysis is carried out at the pixel level; the analysis level and analysis accuracy are difficult to support complex ecological environmental conditions and changes, and to provide in-depth guidance for the analysis of habitat factors at the meso-micro level.

[0008] In guiding ecological construction practice, existing analysis often uses medium- and low-resolution remote sensing images with an accuracy of less than 30M, and the analysis scale is mostly regional and above spatial scales. In medium- and micro-actual environmental development and construction projects, this type of medium- and low-precision analysis has the problem of mixed and vague habitat factors, which makes it difficult to guide actual construction; ecological data acquisition requires manual investigation and annotation, the cost of manual survey is high, and the data update speed is difficult to cope with the changes and needs of high-speed construction; the data units of the research objects at this scale are mostly pixels distributed in a grid point shape, which are abstract data points in the actual space, and lack specific points for the habitat factors and objects contained in the ecological space; the analysis process is mostly calculated and processed based on existing models, and there is a lack of ways to supplement data in response to dynamic changes. The overall dynamic process of the "data collection-processing-analysis" link needs further integration and development. Summary of the invention

[0009] In view of the deficiencies of the prior art, the present invention proposes a quantitative analysis method for surface habitat factors.

[0010] The purpose of the present invention can be achieved through the following technical solutions:

[0011] A quantitative analysis method for surface habitat factors comprises the following steps:

[0012] Step 1: Obtain surface satellite images of the target area;

[0013] Step 2: Classify the surface landscape elements of the target area;

[0014] Step 3: Test and select segmentation scale parameters based on typical habitat factor object types in the base area to classify habitat factor objects;

[0015] Step 4: Extract the band value of each habitat factor object,

[0016] Step 5: Based on the habitat factor parameters obtained in step 3, construct a data set and obtain a high-confidence sample library;

[0017] Step 6: Use the maximum likelihood model to recursively select the optimal features for high-confidence data samples.

[0018] Optionally, after obtaining the satellite image in step 1, the satellite image is preprocessed:

[0019] Step 1.1: Perform radiometric calibration on the raw image data to eliminate the error of the sensor itself;

[0020] Step 1.2: Perform atmospheric correction to eliminate the images reflected by surface landscape elements caused by atmospheric refraction and scattering;

[0021] Step 1.3: Perform orthorectification according to the transformation coefficient matrix parameters;

[0022] Step 1.4: Fusion process the orthorectified remote sensing images and select appropriate remote sensing images for analysis according to the experimental requirements.

[0023] Optionally, in step 3, the habitat factor object is segmented based on a multi-scale segmentation function model, and the parameters of the segmentation function model are configured as one or more of segmentation scale, segmentation layer number, color weight, shape weight, smoothness, and compactness.

[0024] Optionally, the band value in step 4 is one or more of Red band value, Green band value, Blue band value, Nir band value, BRIGHT band value, NDVI band value, NDBI band value, MNDWI band value, and GLCM band value.

[0025] Optionally, step 5 includes the following steps:

[0026] Obtain the maximum value, minimum value, average value and standard deviation of each band value in step 4;

[0027] Select a machine learning multidimensional data sample library based on known habitat factor types;

[0028] The principal component analysis method is used to reduce the dimension of the multidimensional data sample library database of habitat factor machine learning; after multiple training iterations, the high-impact parameters of each type of habitat factor are obtained to build a high-confidence sample library;

[0029] Optionally, the landscape elements are divided into one or more of vegetation elements, water elements, soil elements, and built elements.

[0030] 1 The present invention uses multi-source high-resolution remote sensing image fusion to extract habitat factors, realizing high-efficiency, high-dimensional, and high-refresh-rate habitat factor data collection under various site conditions.

[0031] 2 The present invention utilizes object-oriented intelligent classification technology to upgrade the habitat factor analysis unit from the previous point (pixel) level to the ecological object level with ecological parameter attributes, reducing the salt and pepper phenomenon existing in the previous analysis, and using the habitat factor object as the data set to realize the overall analysis of the multi-dimensional data of the habitat factor object.

[0032] 3 The present invention reduces the low-correlation parameters of habitat factor analysis through dimensionality reduction by principal component analysis, constructs a habitat factor dimensionality reduction parameter database, and improves the speed and efficiency of habitat sample plot image analysis.

[0033] The present invention proposes a method for constructing high-confidence sample data of habitat factors, analyzes the spectral parameters of various habitat factors, reduces the low-correlation parameters through dimensionality reduction by principal component analysis, constructs a high-confidence parameter database of habitat factors, reduces the interference of multidimensional parameters of remote sensing images, improves the speed and efficiency of high-resolution remote sensing image analysis, and enables habitat factors with similar spectra to be distinguished.

[0034] 5 The present invention proposes a maximum likelihood method recursive feature semi-supervised classification learning and analysis method, combined with high confidence sample library learning, to achieve adaptive generalization analysis of habitat factor samples with fewer sample labels, and to achieve high efficiency and high accuracy analysis of multiple habitat factors under complex parameter conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The present invention will be further described below in conjunction with the accompanying drawings.

[0036] Figure 1 Multispectral images of target areas for some examples of this application;

[0037] Figure 2 A panchromatic image of a target area as an example of the present application;

[0038] Figure 3 An orthophoto of a target area as an example of this application;

[0039] Figure 4 A fused image of a target area as an example of the present application;

[0040] Figure 5 A remote sensing image of the target area and its surrounding area as an example of this application;

[0041] Figure 6 A segmentation layer of surface landscape elements of a target area for an example of this application;

[0042] Figure 7 A segmentation layer of surface landscape elements of a target area for an example of this application;

[0043] Figure 8 A segmentation layer of surface landscape elements of a target area for an example of this application;

[0044] Fig. 9 An NDVI interpretation map of surface landscape elements of the target area as an example of this application;

[0045] Fig.10 An NDBI interpretation map of the surface landscape elements of the target area as an example of this application;

[0046] Fig.11 The MNDWI interpretation map of surface landscape elements is an example of this application;

[0047] Fig.12 This is a GLCM interpretation map of surface landscape elements as an example of this application.

[0048] Fig.13 The land cover spectral information of the surface landscape elements of the remote sensing image of the target area as an example of this application;

[0049] Fig.14 This is a flowchart of PCA dimensionality reduction recursive classification of surface habitat factors for an example of this application;

[0050] Fig.15 This is a graph of the classification and analysis results of surface habitat factors in an example of this application. DETAILED DESCRIPTION

[0051] The specific technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0052] The method of the present invention includes seven steps for habitat factor analysis, namely, multi-source remote sensing image preprocessing, habitat factor object classification, multi-layer and multi-scale parameter adjustment clustering of habitat factor objects, habitat factor object library parameter analysis, habitat factor object library parameter establishment, habitat factor analysis parameter PCA dimensionality reduction, and high confidence database sample iterative classification.

[0053] 1) Multi-source remote sensing image preprocessing

[0054] In the multi-source remote sensing image preprocessing step, the original high-resolution remote sensing image processed by the method of the present invention comes from satellite remote sensing images or drone remote sensing images. The present invention uses high-resolution satellite multispectral remote sensing images with a spatial resolution of 3.2 meters and four bands, namely, blue band (0.45-0.52μm), green band (0.52-0.59μm), red band (0.63-0.69μm), and near infrared band (0.77-0.89μm). In some specific examples of the present invention, the processed remote sensing images are as follows Figure 1 , Figure 2shown.

[0055] First, the original image data is radiometrically calibrated to eliminate the error of the sensor itself. Radiometric calibration converts the original pixel digital quantization value (Digital Number, DN) of the remote sensing image into atmospheric reflectivity to obtain the real reflection data of the surface landscape elements collected by the sensor. According to actual needs, it can be divided into absolute calibration and relative calibration. This example uses absolute calibration, and its conversion formula is as follows:

[0056] Lλ=Gain*DN+Offset

[0057] Where: Lλ is the radiation brightness value, Gain is the gain of each band, Offset is the offset of each band, and the gain and offset information are obtained from the header file of the remote sensing image.

[0058] Secondly, atmospheric correction is performed to eliminate the images reflected by surface landscape elements due to atmospheric refraction, scattering and other phenomena. In the example of the present invention, the FLAASH atmospheric correction model and the improved MODTRAN radiation transfer algorithm are used to invert the atmospheric properties of the imaging area by using the features of the remote sensing image pixel spectrum to replace the measured atmospheric parameters during image imaging, eliminate the scattering effects of aerosols and water vapor, and obtain a surface reflectivity with high accuracy. In the subsequent analysis steps, good universality and high accuracy can be achieved. After being processed by the FLAASH atmospheric correction model, a base remote sensing image with actual surface reflectivity is obtained, such as Figure 3 shown.

[0059] At the same time, according to the RPC (rational polynomial coefficients, transformation coefficient matrix) parameters, orthorectification is performed and the projection is WGS84 / UTM. Remote sensing orthorectification selects ground control points and combines the digital elevation model (DEM) data within the image range to perform projection error correction and tilt correction, correct the pixel offset caused by terrain changes in the remote sensing image, and resample the remote sensing image into an orthophoto. Figure 4 shown.

[0060] Finally, the orthorectified remote sensing images are fused and the appropriate remote sensing images are selected for analysis according to the experimental requirements. In this example, panchromatic band (2.1m) and multispectral band (5.8m) remote sensing images are used. In order to improve the accuracy and feature recognition ability of remote sensing images and enhance the availability of remote sensing image data, the panchromatic band (2.1m) and multispectral band (5.8m) remote sensing images are fused. The use of HIS transformation fusion can improve the spatial resolution of the image, better preserve the spectral characteristics of the surface landscape elements, and improve the accuracy and interpretability of the remote sensing image. The spatial resolution of the ZY-3 multispectral image after HIS transformation fusion is 2.1m, and the bands are blue band (Blue), green band (Green), red band (Red) and near infrared band (NIR). In order to improve the interpretation speed of surface landscape elements and reduce redundant calculations, the base and its surrounding parts of the fused image are intercepted for classification processing, such as Figure 5 shown.

[0061] 2) Classification of base habitat factors

[0062] According to the field survey, combined with the characteristics of remote sensing image interpretation technology and the performance of remote sensing images of surface landscape elements, the surface landscape elements of the Songlingtou target area of ​​the base were first divided into four categories: vegetation elements, water elements, soil elements, and built elements. On this basis, all surface ecological sensitive elements were divided into 10 categories: trees, shrubs, ground cover, tidal flats, roads, bare land, water bodies, tea gardens, bamboo forests, and buildings. (Table 1)

[0063] Table 1 Classification of ecological sensitive elements on the surface of Songlingtou area

[0064]

[0065] 3) Clustering of multi-layer and multi-scale parameter adjustment of habitat factors

[0066] Based on the typical habitat factor object types in the base area, the segmentation scale parameters are tested and selected to classify the habitat factor objects.

[0067] This case adopts a multi-resolution segmentation model, using segmentation scale, segmentation layer number, color weight, shape weight, smoothness, and compactness to complete the segmentation of habitat factor objects. The determination of segmentation rules mainly considers the practical segmentation of images based on the spectral characteristics, geometric shapes, and topological characteristics of the habitat factor object elements, and forms a multi-scale remote sensing image object hierarchical network with three levels of a, b, and c. The parameters of each layer are as follows.

[0068] The segmentation parameters of layer a are: scale 100, color weight 0.9, shape weight 0.4, smoothness 0.5, and compactness 0.5; the segmentation parameters of layer b are: scale 50, color weight 0.5, shape weight 0.5, smoothness 0.3, and compactness 0.3; the segmentation parameters of layer c are: scale 30, color weight 0.9, shape weight 0.1, smoothness 0.5, and compactness 0.2. According to multiple experimental analyses, the above parameter combination can effectively segment the habitat factors in the test site.

[0069] The segmentation scale should not be too large to avoid fragmentation of the ground features. Considering the high degree of interlacing between water bodies and non-water bodies in Songlingtou area, it is advisable to distinguish water bodies from non-water bodies at a macro scale, and its segmentation scale is adjusted to 100; Layer b mainly distinguishes between trees, shrubs, ground cover, bare land, and tidal flats. On the basis of inheriting the segmentation of layer a, combined with the shape and texture characteristics of trees, shrubs, ground cover, bare land, tidal flats and other natural evolution characteristics, the segmentation scale parameter is 50 to separate the natural elements; Layer c mainly distinguishes bamboo forests, tea gardens, buildings, roads and other surface landscape elements with artificial intervention characteristics. By analyzing the spectrum, shape, compactness, smoothness and other characteristics presented by them, and comparing them with the existing site basic data, the agricultural vegetation and built landscape information elements are effectively extracted, such as Figures 6 to 8 Table 2 shows the scale parameters and discrimination index of each level segmentation.

[0070] Table 2: Multi-layer segmentation scale and discrimination parameters

[0071]

[0072] 4) Parameter analysis of habitat factor object library

[0073] The parameters of the habitat factor object obtained in step 2) are parsed to obtain their parameters respectively.

[0074] Extract the Red band value of each habitat factor object. The calculation formula is as follows:

[0075]

[0076] Where x is the current calculated habitat factor object, i is a pixel in the current habitat factor object x, n is the set of all pixels in the current calculated habitat factor object x, K is the red band value of a pixel i in the current calculation sub-superpixel x. x (red) is the Red band value of the currently calculated habitat factor object x.

[0077] Extract the Green band value of each habitat factor object. The calculation formula is as follows:

[0078]

[0079] Where x is the current calculated habitat factor object, i is a pixel in the current habitat factor object x, n is the set of all pixels in the current calculated habitat factor object x, K is the red band value of a pixel i in the current habitat factor object x. x (Green) is the Green band value of the currently calculated habitat factor object x.

[0080] Extract the Blue band value of each habitat factor object. The calculation formula is as follows:

[0081]

[0082] Where x is the current calculated habitat factor object, i is a pixel in the current habitat factor object x, n is the set of all pixels in the current calculated habitat factor object x, K is the red band value of a pixel i in the current habitat factor object x. x (Blue) is the Blue band value of the currently calculated habitat factor object x.

[0083] Extract the Nir band value of each habitat factor object. The calculation formula is as follows:

[0084]

[0085] Where x is the current calculated habitat factor object, i is a pixel in the current habitat factor object x, n is the set of all pixels in the current calculated habitat factor object x, K is the red band value of a pixel i in the current habitat factor object x. x (Nir) is the Nir band value of the currently calculated habitat factor object x.

[0086] Extract the BRIGHT band value of each habitat factor object. The calculation formula is as follows:

[0087]

[0088] Where x is the current calculated habitat factor object, i is a pixel in the current habitat factor object x, n is the set of all pixels in the current calculated habitat factor object x, K is the red band value of a pixel i in the current habitat factor object x. x (Bright) is the Bright band value of the currently calculated habitat factor object x.

[0089] Calculate the normalized difference water index mndwi for each habitat factor object using the following formula:

[0090]

[0091] Among them, x is the current calculation habitat factor object, i is a pixel point in the current calculation habitat factor object x, n(x) is the set of all pixel points in the current calculation habitat factor object x, is the green band value of a pixel i in the current habitat factor object x. is the Mir band value of a pixel i in the current habitat factor object x, K x (mndwi) is the normalized difference water index value of the current calculated habitat factor object x, -1<=K x (mndwi)<=1, positive value indicates water area, otherwise it indicates land area.

[0092] Calculate the normalized vegetation index ndvi for each habitat factor object using the following formula:

[0093] The normalized vegetation index K of the habitat factor object x (ndvi) is calculated as follows:

[0094]

[0095] Among them, x is the current calculation habitat factor object, i is a pixel point in the current calculation habitat factor object x, and n is the set of all pixel points in the current calculation habitat factor object x. is the red band value of a pixel i in the current habitat factor object x. is the near-infrared Nir band value of a pixel i in the current habitat factor object x, K x (ndvi) is the normalized vegetation index value of the current calculated habitat factor object x, such as Fig. 9 shown.

[0096] Calculate the normalized building index ndbi for each habitat factor object using the following formula:

[0097]

[0098] Among them, x is the current calculation habitat factor object, i is a pixel point in the current calculation habitat factor object x, and n is the set of all pixel points in the current calculation habitat factor object x. is the red band value of a pixel i in the current habitat factor object x. is the near-infrared Nir band value of a pixel i in the current habitat factor object x, Kx (ndvi) is the normalized vegetation index value of the current calculated habitat factor object x, such as Fig.10 shown.

[0099] Calculate the normalized difference water index mndwi for each habitat factor object using the following formula:

[0100]

[0101] Among them, x is the current calculation habitat factor object, i is a pixel in the current calculation habitat factor object x, n(x) is the set of all pixels in the current calculation habitat factor object x, is the green band value of a pixel i in the current habitat factor object x. is the Mir band value of a pixel i in the current habitat factor object x, K x (mndwi) is the normalized difference water index value of the current calculated habitat factor object x, -1<=K x (mndwi) <= 1, positive value indicates water area, otherwise it indicates land area, such as Fig.11 shown.

[0102] 5) Establishment of parameters of habitat factor object library

[0103] According to the habitat factor parameters obtained in step 3), a data set is constructed.

[0104] A habitat factor database is constructed according to the habitat factor types within the site. The habitat factor parameters in the study area are analyzed to obtain the maximum, minimum, mean, and standard deviation of the Red band of each habitat factor object; the maximum, minimum, mean, and standard deviation of the Green band; the maximum, minimum, mean, and standard deviation of the Blue band; the maximum, minimum, mean, and standard deviation of the Nir band; the mean of the BRIGHT band; the maximum, minimum, mean, and standard deviation of the Normalized Difference Vegetation Index (NDVI); the maximum, minimum, mean, and standard deviation of the Normalized Difference Water Index (MNDWI); the maximum, minimum, mean, and standard deviation of the Normalized Building Index (NDBI); the maximum, minimum, mean, and standard deviation of the Gray Level Co-occurrence Matrix (GLCM); the object morphological parameters: compactness value; aspect ratio value, texture feature value. The interpretation diagram is as follows Fig.12 shown.

[0105] A machine learning multidimensional data sample library is selected based on known habitat factor types.

[0106] Based on the fused multispectral remote sensing images, combined with the surface habitat factor data of Songlingtou area obtained by manual on-site investigation, the habitat factors in the site were further corrected using the visual texture characteristics of the human eye and prior knowledge. Ten types of habitat factor elements in Songlingtou area of ​​the test site were selected, including trees, shrubs, ground cover, bamboo forest, tea garden, water body, building, road, bare land, and tidal flat. The number of objects in each type ranged from 50 to 200, and a multidimensional data sample library of habitat factor machine learning was constructed. The sample library data standards are as follows: Fig.13 shown.

[0107] 6) PCA dimensionality reduction of habitat factor analysis parameters

[0108] The principal component analysis method is used to reduce the dimension of the multidimensional data sample library database of habitat factor machine learning. After multiple training iterations, the high-impact parameters of each type of habitat factor are obtained to build a high-confidence sample library.

[0109] The maximum likelihood model is used to recursively select the optimal features for high confidence data samples.

[0110] First, configure the analysis bands: Red band, Green band, Blue band, Nir band, NDVI band, NDBI band, MNDWI band, and GLCM band.

[0111] Among them, the principal component analysis (PCA) method is used to reduce the dimension of the multidimensional data sample library database of habitat factors based on eigenvalue analysis. The specific method can adopt the following steps.

[0112] (1) Input a sample set of multidimensional data of habitat factors and determine the number of spatial dimensions to be reduced.

[0113] Input habitat factor multidimensional data sample set D i ,

[0114] D i ={x1,x2,x3,…,x m}

[0115] Determine the dimension d' of the low-dimensional space of multidimensional data of habitat factors. After parameter selection experiments, d' takes the minimum value that makes the following formula valid when t = 95%.

[0116]

[0117] (2) Standardize the matrix data and calculate the mean and standard deviation of each variable in the multidimensional data sample of habitat factors.

[0118] Centralize all multidimensional data samples of habitat factors:

[0119]

[0120] (3) Correlation coefficient matrix function, calculate eigenvectors and eigenvalues.

[0121] Calculate the covariance matrix XX of multidimensional data samples of all habitat factors T , the covariance matrix of the multidimensional data sample of the habitat factor XX T Do eigenvalue decomposition.

[0122] Take the eigenvectors w1,w2,w3,…,w corresponding to the eigenvalues ​​of the low-dimensional space dimension d′ d′ .

[0123] (4) Output the projection matrix W of the multidimensional data sample of the habitat factor i

[0124] W i ={w1,w2,w3,…,w d′}

[0125] According to the multidimensional data sample projection matrix W of the habitat factor i , construct the influencing parameters of multiple categories of habitat factors, and build a high-confidence sample library of multidimensional data of habitat factors.

[0126] 7) MLE iterative classification of high confidence database samples

[0127] The maximum likelihood estimation (MLE) method was used to recursively select the optimal features of the habitat factor data samples after dimensionality reduction.

[0128] First, the reduced-dimension dataset of habitat factors was preprocessed. Training cases were extracted from the habitat factor dataset based on the expert prior recognition and field sampling survey results. The reduced-dimension dataset of habitat factors was divided into labeled data and unlabeled data. The labeled data was configured with ten types of habitat factors: trees, shrubs, ground cover, bamboo forest, tea garden, water body, building, road, bare land, and tidal flat. The unlabeled data was mixed with high-confidence samples to prepare for machine learning by the recursive classifier.

[0129] Secondly, set the iteration conditions of the recursive classifier, the output result category is: 10, the elimination score is: 0.01, and the prior probability weight is: EQUAL. When the input data does not meet the conditions, it will be used in the high confidence sample library for further iterative training; when the input data meets the conditions, the primary classifier is output.

[0130] Then, the accuracy of the classification analysis results of the primary classifier is verified, and the confusion matrix algorithm (Error Matrix Based on Samples, EMBS) is used to establish the error matrix for measurement evaluation and regression analysis. The overall accuracy (OA) and Kappa coefficient are selected as two measurement evaluation indicators (producer accuracy (PA) and user accuracy (UA) are used for reference). The relevant algorithm is expected to be as follows:

[0131]

[0132] Where: OA is the overall classification accuracy, r is the number of rows or columns of the error matrix, x ii is the number of a certain category on the main diagonal of the matrix, and N is the total number of reference samples.

[0133]

[0134] Where K is the Kappa coefficient, r is the number of rows in the error matrix, and x ii is the value of the i-th row and i-th column on the diagonal (i.e., the number of correct classifications of a certain type of habitat factor), x i+ and x +i are the sum of the i-th row and i-th column respectively, and N is the total number of reference samples.

[0135] Iteratively upgrade the primary classifier, set the ideal expected classification accuracy OA ≥ 90%, Kappa ≥ 85%, if the generalized classification result of the primary classifier does not meet the accuracy verification conditions, then perform classifier optimization, that is, the recursive classifier restarts the iteration until the generalized classification accuracy of the primary classifier reaches the accuracy verification conditions; if the generalized classification result of the primary classifier meets the accuracy verification conditions, then output the final classifier of the habitat factor, such as Fig.14 shown.

[0136] Finally, the final classifier of habitat factors is used to generalize and classify the habitat factors of the experimental plots, and the classification and analysis results are output, such as Fig.15 shown.

[0137] In the description of this specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0138] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements all fall within the scope of the present invention to be protected.

Claims

1. A method for analyzing surface habitat factors, characterized in that: The following steps are involved: Step 1: Obtain surface satellite images of the target area; Step 2: Classify the surface landscape elements of the target area; Step 3: Test and select the segmentation scale parameters according to the typical habitat factor object types in the base area to classify the habitat factor objects; Step 4: extracting the band value of each habitat factor object; Step 5: Based on the habitat factor parameters obtained in step 3, construct a data set and obtain a high-confidence sample library; Step 6: Use the maximum likelihood model to recursively select the optimal features for high confidence data samples; After the satellite image is obtained in step 1, the satellite image is preprocessed: Step 1.1: Perform radiometric calibration on the raw image data to eliminate the error of the sensor itself; Step 1.2: Perform atmospheric correction to eliminate the images reflected by surface landscape elements caused by atmospheric refraction and scattering; Step 1.3: Perform orthorectification according to the transformation coefficient matrix parameters; Step 1.4: Fusion process the orthorectified remote sensing images and select appropriate remote sensing images for analysis according to the experimental requirements; The step 3 specifically includes: using a multi-scale segmentation function model, using a total of 6 parameters including segmentation scale, segmentation layer number, color weight, shape weight, smoothness, and compactness to complete the segmentation of habitat factor objects. The determination of the segmentation rules mainly considers the practical segmentation of the image according to the spectral characteristics, geometric shape, and topological characteristics of the habitat factor object elements, and forms a multi-scale remote sensing image object hierarchy network with three levels of a, b, and c, wherein layer a is for distinguishing water bodies and non-water bodies from a macroscopic scale, layer b is for the shape and texture characteristics of trees, shrubs, ground covers, bare land, and tidal flats with natural evolution characteristics, and layer c is for bamboo forests, tea gardens, buildings, roads, and other surface landscape elements with artificial intervention characteristics. By analyzing the spectrum, shape, compactness, and smoothness presented by them, and comparing them with the existing site basic data, natural elements, agricultural vegetation, and built landscape information elements can be effectively extracted.

2. The method for analyzing surface habitat factors according to claim 1, characterized in that: In the step 3, the habitat factor object is segmented based on a multi-scale segmentation function model, and the parameters of the segmentation function model are configured as one or more of segmentation scale, number of segmentation layers, color weight, shape weight, smoothness, and compactness.

3. The method for analyzing surface habitat factors according to claim 1, characterized in that: The band value in step 4 is one or more of Red band value, Green band value, Blue band value, Nir band value, BRIGHT band value, NDVI band value, NDBI band value, MNDWI band value, and GLCM band value.

4. The method for analyzing surface habitat factors according to claim 1, characterized in that: The step 5 comprises the following steps: Obtain the maximum value, minimum value, average value and standard deviation of each band value in step 4; Select a machine learning multidimensional data sample library based on known habitat factor types; The principal component analysis method is used to reduce the dimensionality of the data in the multidimensional data sample library of habitat factor machine learning; after multiple training iterations, the high-influence parameters of each type of habitat factor are obtained to construct a high-confidence sample library.

5. The method for analyzing surface habitat factors according to claim 1, characterized in that: The landscape elements are divided into one or more of vegetation elements, water elements, soil elements, and built elements.