CN114462512A - Systematic hyperspectral grassland community division method - Google Patents

Systematic hyperspectral grassland community division method Download PDF

Info

Publication number
CN114462512A
CN114462512A CN202210053671.9A CN202210053671A CN114462512A CN 114462512 A CN114462512 A CN 114462512A CN 202210053671 A CN202210053671 A CN 202210053671A CN 114462512 A CN114462512 A CN 114462512A
Authority
CN
China
Prior art keywords
data
hyperspectral
area
image
grassland
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202210053671.9A
Other languages
Chinese (zh)
Other versions
CN114462512B (en
Inventor
魏丹丹
刘凯
肖晨超
黄熙枝
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Ministry Of Natural Resources Land Satellite Remote Sensing Application Center
Original Assignee
Ministry Of Natural Resources Land Satellite Remote Sensing Application Center
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Ministry Of Natural Resources Land Satellite Remote Sensing Application Center filed Critical Ministry Of Natural Resources Land Satellite Remote Sensing Application Center
Priority to CN202210053671.9A priority Critical patent/CN114462512B/en
Publication of CN114462512A publication Critical patent/CN114462512A/en
Application granted granted Critical
Publication of CN114462512B publication Critical patent/CN114462512B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features

Landscapes

  • Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Artificial Intelligence (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Image Processing (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a systematic hyperspectral grassland community dividing method, which relates to the field of remote sensing image application and comprises the following steps: step 1, selecting sample label data by combining a plurality of data; 2, selecting a proper waveband subset by using a linear prediction method and constructing a spectral feature; step 3, extracting spatial features of the selected wave bands by using an extended morphology method; step 4, stacking the spectral features and the spatial features, fusing in a low-dimensional space by utilizing principal component analysis, and obtaining a pre-classification result by utilizing a Random Forest (RF); and 5, classifying and post-processing the classification result by using a label similarity probability filter, and further classifying by using RF (radio frequency) to obtain a final grassland classification result. The method fully utilizes the space-spectrum characteristics of the hyperspectral image, solves the problem of dimension disaster in characteristic extraction, and relieves the influence of noise on classification results through a classification post-processing method, thereby improving the identification precision of the grassland community.

Description

Systematic hyperspectral grassland community division method
Technical Field
The invention relates to the field of remote sensing image application, in particular to a systematic hyperspectral grassland community division method.
Background
The grassland degradation is mainly characterized in that the structure of a vegetation community is changed, and the grassland community identification and division based on hyperspectrum are the basis and the premise of monitoring and treating the degradation of the grassland with large area and high precision by utilizing remote sensing. The method has the advantages that the method is high in mixing degree and low in vegetation, multispectral remote sensing identification and division difficulty is high, each pixel in the hyperspectral remote sensing image is composed of dozens of narrow wave bands or even hundreds of narrow wave bands, and the method can be used for fine grassland community division through classification operation.
The hyperspectral images contain abundant spatial information and spectral information, effective feature extraction is a key step for realizing the fine classification of the hyperspectral images, otherwise, a large amount of computing resources are wasted, and the hyperspectral images have low classification precision. In the conventional method, only spectral features or spatial features are generally utilized, but the contribution of the corresponding spatial features or spectral features to classification is ignored, and in addition, the method of stacking and fusing the space-spectrum features has the problem of dimension disaster of hyperspectral data. In recent years, deep learning methods represented by convolutional neural networks are widely applied to hyperspectral image classification in an end-to-end training mode, but the application of the deep learning methods in actual production is limited due to the lack of interpretability and the need of consuming a large amount of manpower and material resources for parameter adjustment. Therefore, the invention provides a systematic operation flow from the sample label data acquisition to the feature extraction and fusion to the classification result post-processing, and can be applied in specific scenes.
Disclosure of Invention
The invention provides a systematic hyperspectral grassland community division method aiming at the problems of insufficient utilization of characteristics, long training time of a deep learning method, difficult parameter adjustment and the like of the traditional method in hyperspectral grassland community division, and is used for improving the characteristic utilization rate and the classification precision.
The systematic hyperspectral grassland community division method comprises the following steps:
step S1, selecting partial pixels in the research area range as sample label data by utilizing actually measured sample point data, the vegetation map (1: 1000000) data (vegetation map data for short) and the national-level natural protection area functional division data (protection area data for short);
s2, inputting a hyperspectral image, selecting a proper wave band subset by using a Linear Prediction (LP) based method, and extracting spectral features;
s3, obtaining the spatial characteristics of the hyperspectral image by the selected wave band by using an Extended Morphology (EMP) method;
s4, stacking the extracted spectral features and spatial features, fusing in a low-dimensional space by using a Principal Component Analysis (PCA), and obtaining a pre-classification result by combining sample label data and using a Random Forest (RF);
and step S5, classifying and post-processing the classification result by using a Label Similarity Probability Filter (LSPF) method, and further classifying by using RF (radio frequency) to reduce the influence of salt and pepper noise on the classification result.
The invention is also characterized in that:
when step S1 is executed, sample label data is selected mainly by actually measuring sample point data, vegetation map data, and protected area data, in combination with image data, and the specific process is as follows:
s1.1, actually measured sample point data is data obtained by manual field investigation, and comprises longitude and latitude coordinates and grassland dominant species information in a 1m x 1m grid, then a ZY1-02D hyperspectral image with the acquisition time similar to that of the actually measured sample point data is selected, a 200m x 200m buffer zone is made according to the actually measured sample point data, and pixels in the buffer zone are used as sample label data;
s1.2, the vegetation map data is published by geological publishing houses in 2008, and the geographical spatial distribution of 11 vegetation type groups, more than 900 groups and sub-groups of 55 vegetation types and more than about 2000 groups dominant species in China is included; the protection area data is a natural centralized distribution area of a country for a representative natural ecosystem and rare or endangered wild animal and plant species, and comprises a core area, a buffer area and an experimental area, wherein the core area is a core area for protecting the ecosystem, is forbidden for any unit and person to enter, and is not allowed to conduct scientific research activities, the periphery of the core area can be divided into buffer areas with certain areas, only scientific research observation activities are conducted, the periphery of the buffer areas is divided into the experimental area, and activities such as scientific experiments, teaching practice, visiting investigation, traveling, breeding rare or endangered wild animals and plants can be conducted; overlaying vegetation map data, protection area data and image data in ArcGIS, wherein communities in the core area and the buffer area are considered to be unchanged for many years, but communities in the experimental area need to combine with the current image to select a sample label;
and S1.3, under the precondition that vegetation map data, protection area data and image data are superposed, regarding a special area, such as a built-up area (a non-agricultural production construction area developed by land acquisition and actual construction within a municipal administrative district), regarding the municipal district as a main part and referring to the vegetation map data, and regarding a community suitable for growing in the built-up area as a community corresponding to the municipal district.
When step S2 is executed, the method mainly includes performing band selection on the hyperspectral image and extracting spectral features, and includes the specific steps of:
s2.1, selecting a group of wave bands B in the hyperspectral image1And B2Initializing the algorithm and then generating the band subset Φ ═ B1B2}；
Step S2.2, selecting B which is the most different from the current phi from all wave bands by the methods of estimation and linear prediction3And updating the band subset Φ ═ Φ U { B ═ Φ U }3}；
a0+a1B1+a2B2＝B′ (1)
Wherein B' is B3Estimation and linear prediction of results of (a)0，a1，a2Is to make the linearity error e | | | B3-the minimum parameter of B' |;
step S2.3, repeating the step S2.2 until phi reaches the preset wave band subset number, and obtaining the spectral characteristic yspe＝Φ。
When step S3 is executed, the method mainly includes performing on-off transformation on the hyperspectral image after the band selection, and the specific process is as follows:
s3.1, performing corrosion-first expansion transformation and then expansion transformation on the hyperspectral pixel x, wherein the operation is defined as follows:
Figure BDA0003475344880000031
wherein, MPγRepresenting the on-transformed image; gamma rayk(x) Representing an open operation performed on image x; l is the number of opening operations;
s3.2, performing expansion-first and corrosion-second closed transformation on the hyperspectral pixel x, wherein the operation is defined as follows:
Figure BDA0003475344880000032
wherein, MPΦRepresenting the closed transformed image; phik(x) Indicating a closing operation performed on the image x; p is the number of closed operations; in the formulas (2) and (3), k is the size of the structural element, and when k is 0, it indicates that the original image is not operated;
s3.3, combining the original hyperspectral pixel with the result after the opening and closing operation to obtain an EMP result, namely the spatial characteristic yspa
yspa＝EMP＝{MPγ，x，MPΦ} (4)
When step S4 is executed, the method mainly includes stacking spectral features and spatial features of the hyperspectral image, and fusing in a low-dimensional space by using PCA, so as to classify by using RF, which includes the specific steps of:
s4.1, stacking and normalizing the spectral features and the spatial features in the steps S2 and S3 to obtain fused features
Figure BDA0003475344880000033
Step S4.2, data
Figure BDA0003475344880000034
From n using PCAThe dimensional space is projected into the m-dimensional space:
Figure BDA0003475344880000035
wherein, P ═ P1，...，Pm]And (m < n) represents a transformation matrix formed by selecting eigenvectors corresponding to the first m eigenvalues (namely m-dimensional space) according to the eigenvalue arrangement sequence.
And S4.3, inputting the obtained space-spectrum fusion characteristic G and the sample label data into an RF classifier for training, and testing the image by using the model obtained by training to obtain a grassland community pre-classification result.
When step S5 is executed, the obtained pre-classification result is subjected to post-classification processing by using LSPF (Label Similarity Probability Filter), and then a grassland community classification result is obtained by using an RF classifier, so as to solve salt and pepper noise caused by spectrum Similarity.
And S5.1, according to whether the result pixel belongs to the category c or not, mapping and converting the pre-classification result into a series of binary label images, wherein the binary label images are set to be 1 if the result pixel belongs to the category c, and otherwise, the binary label images are 0, for example, if 5 types of ground objects exist in the image scene, the pre-classification result images are divided into 5 layers of binary label images.
Step S5.2, within each spatial window, the cumulative probability of pixel (i, j) being of class c is calculated as fc(i, j), then traversing all classes to obtain the cumulative probability data of all pixels as
Figure BDA0003475344880000036
Can be expressed as a 1 × C vector, then the cumulative probability of all pixels and the sample label data are input into the RF classifier, and F is determined by using the trained modelcAnd (i, j) testing to obtain a classification result graph of the high spectral data of the grassland community.
According to the method, spectral features and spatial features extracted from the hyperspectral remote sensing images are fused in a low-dimensional space by using a PCA (principal component analysis) method, so that space-spectrum features of the hyperspectral images are mined, the problem of dimensionality disaster of the hyperspectral images is solved, then the fused features are classified by using RF (radio frequency) to obtain a pre-classification result, then the classified features are classified by using LSPF (least squares filter function) to process the classified features so as to solve the phenomenon of salt and pepper noise in the classification result, and finally the RF is used for obtaining a final classification result. The method can effectively solve the problems of poor classification performance, low precision and the like caused by spectral characteristics or spatial characteristics, and provides a set of complete classification flow for identifying the grassland community.
Drawings
FIG. 1 is a flow chart of an embodiment of the present invention.
Detailed Description
In order to make the technical solutions and technical problems of the present invention more clearly understood, the present invention is further described with reference to fig. 1 as an embodiment.
The invention provides a systematic hyperspectral grassland community division method, and the grassland is taken as an important component of an ecosystem and plays an important role in protecting water and soil, preventing wind and fixing sand and protecting biological diversity, but in recent years, due to the influence of climate change and human activities, large-area grassland degradation has already occurred in China, so that the monitoring of the grassland degradation is particularly important. The method has wide application in grassland degradation monitoring by utilizing hyperspectral remote sensing, but further application is limited due to the problems of insufficient feature extraction and 'dimensionality disaster', and challenges are brought to large-area grassland degradation monitoring.
S1, selecting sample label data determines the accuracy of grassland community division to a certain extent, therefore, in order to ensure the reasonability of sample selection, the invention takes the Sinkiang Neugilan of inner Mongolia as a research area, and combines the data of actually measured sample points, the vegetation map (1: 1000000) data (vegetation map data for short), the functional data of national natural protection area (protection area data for short) and ZY1-02D hyperspectral data to jointly construct sample label data, and the specific process comprises the following steps:
s1.1, carrying out field investigation in 7 and 8 months of luxuriant grassland vegetation, recording longitude and latitude coordinates of investigation points and grassland dominant vegetation types, selecting 4 inner Mongolia Silo-Lei union ZY1-02D hyperspectral images with cloud cover less than 20% and imaging time of 7 and 8 months, making a buffer area of 200m x 200m around actual measurement points, and taking pixels in the buffer area as sample label data;
s1.2, the vegetation map data is published by geological publishing houses in 2008, and the vegetation map data comprises 11 vegetation type groups in China, the geographical spatial distribution of 900 more than 900 groups and sub-groups of 55 vegetation types and about 2000 more than community dominant species, and the vegetation map data covering the Sirina Guo union is selected; the protection area data is a natural centralized distribution area of a country for a representative natural ecosystem and rare or endangered wild animal and plant species, and comprises a core area, a buffer area and an experimental area, wherein the core area is a core area for protecting the ecosystem, is forbidden for any unit and individual to enter, and is not allowed to conduct scientific research activities, the periphery of the core area can be divided into buffer areas with certain areas, only scientific research observation activities are conducted, the periphery of the buffer areas is divided into the experimental area, activities such as scientific experiments, teaching practice, visiting investigation, traveling, breeding rare or endangered wild animals and plants and the like can be conducted, and the Ceylor leau protection area data is selected; overlaying vegetation map data, protection area data and image data in ArcGIS, wherein communities in the core area and the buffer area are considered to be unchanged for many years, but communities in the experimental area need to combine with the current image to select a sample label;
s1.3, under the precondition that vegetation map data, protection area data and image data are superposed, regarding a special area, such as a built-up area (a non-agricultural production construction area developed by land acquisition and actual construction in a municipal administrative area range), regarding a municipal area as a main part and referring to the vegetation map data, regarding a community suitable for growing in the built-up area as a community corresponding to the municipal area;
step S1.4, through the steps S1.1-1.3, 10 types of grassland groups are selected, wherein the types of the grassland groups are respectively stipa glauca prairie (13655), splendid achnatherum salicornia meadow (7775), crops (1384), junglegrass (3822), ulmus macrocarpa forest (2249), ophicumphala (2137), caragana shrub (1266), Betula bacopa (1168), miscellaneous grassland meadow grassland (1354) and jungler cryptophyte (1146), and the number of the selected hyperspectral image pixels in brackets is 35956 in total.
And step S2, selecting a proper waveband subset from the hyperspectral data by using a Linear Prediction (LP) method to form the spectral characteristics of the hyperspectral data.
S2.1, randomly selecting a group of wave bands B from original hyperspectral data1And B2Initializing the algorithm and generating a subset of bands Φ ═ B1B2}；
Step S2.2, the method of estimation and linear prediction allows to evaluate the similarity between the bands, thus allowing to select from all the bands the band B that is the most different from the current phi3And updating the band subset Φ ═ Φ U { B ═ Φ U }3}，
a0+a1B1+a2B2＝B′ (1)
Wherein B' is B3Estimation and linear prediction of results of (a)0，a1，a2Is to make the linearity error e | | | B3-the minimum parameter of B' |;
step S2.3, repeating step S2.2 until the number of phi middle wave band subsets is 20, and enabling yspeThe spectral feature is obtained as phi, and all information of the high-spectrum data can be represented by the least band number as far as possible, so that data redundancy is reduced, and the extraction of the spatial feature is facilitated.
And S3, extracting spatial features from the spectral features phi in the S2 by using an extended morphology method (EMP), and reducing the influence of noise on the classification of the hyperspectral images, wherein the method mainly comprises open transformation and closed transformation.
S3.1, performing corrosion-first expansion transformation on the hyperspectral pixel x, wherein the operation definition is as follows:
Figure BDA0003475344880000051
wherein, MPγRepresenting the on-transformed image; gamma rayk(x) Representing an open operation performed on image x; l is the number of 5-division operations.
S3.2, performing expansion-first and corrosion-second closed transformation on the hyperspectral pixel x, wherein the operation definition is as follows:
Figure BDA0003475344880000052
wherein, MPΦRepresenting the closed transformed image; phik(x) Indicating a closing operation performed on the image x; p is 5, the number of closed operations; in equations (2) and (3), k is the size of the structural element, and when k is 0, it indicates that no operation is performed on the original image.
Step S3.3, based on the morphological features after the "on transform" and the "off transform", selecting the corresponding spectral band to construct the morphological contour, where the morphological contour of the pixel x is mx (2l +1) ═ 220, where m is the number of spectral bands, i.e. y is the number of spectral bandsspa＝EMP
EMP＝{MPγ，x，MPΦ} (4)
Abundant spatial information can be acquired by using an Extended Morphology (EMP) feature extraction method, and structural objects or substances can be distinguished.
And S4, stacking the spectral features and the spatial features of the hyperspectral images to obtain the space-spectral features of the hyperspectral images, and in order to avoid the problem of dimension disaster caused by high-dimensional characteristics, fusing in a low-dimensional space by using a Principal Component Analysis (PCA) method.
S4.1, stacking and normalizing the spectral features and the spatial features in the steps S2 and S3 to obtain fused features
Figure BDA0003475344880000061
The normalization operation can reduce the difference between the data.
Step S4.2, data
Figure BDA0003475344880000062
Projection from n-220 dimensional space to m-50 dimensional space using PCA (m < n):
Figure BDA0003475344880000063
wherein, P ═ P1，...，Pm]And (m < n) represents a transformation matrix composed of eigenvectors corresponding to the first 50 eigenvalues (i.e., m-dimensional space) selected according to the data eigenvalue arrangement order.
And S4.3, inputting the obtained space-spectrum fusion characteristic G and the sample label data into an RF classifier for training, and testing the image by using the model obtained by training to obtain a grassland community pre-classification result. Since the RF is classified based on the pixel characteristics, salt and pepper noise is generated, which causes inaccurate classification of grassland, and thus it is necessary to perform a post-classification processing operation on the pre-classification result.
Step S5, using Label Similarity Probability Filter (LSPF) method to further classify and post-process the classification result, using the spatial correlation with the surrounding pixels to estimate the label probability of all pixels in all classes, then inputting the probability data and sample label data into the classifier to refine the classification diagram obtained in advance, and reducing the influence of salt and pepper noise on the classification result.
And S5.1, mapping and converting the initial classification result into a series of binary label graphs according to whether the pixel belongs to the class c, wherein the class is set to be 1, and if the class is not 0, for example, 10 types of ground objects exist in the image scene, the pre-classification result graph is divided into 10 layers of binary label graphs.
Step S5.2, according to the simplicity and effectiveness of two-dimensional Gaussian distribution in the real world, modeling LSP (Label Similarity Probability) in a local space window by using an exponential function:
Figure BDA0003475344880000064
where g (x, y) is the LSP between the central pixel (with a value of 1) and its neighboring pixels in the same spatial window, e is a natural constant, and σ is the standard deviation.
Step S5.3, further within each spatial window (5 × 5), LSPF (Label Similarity Probability Filter) for each class c can be expressed as:
Figure BDA0003475344880000071
wherein f isc(i, j) is the cumulative probability of the center pixel (i, j) class c, which is a Hadamard product (Hadamard product), Gij＝[g(x，y)]For the LSP matrices in the (i, j) _ th spatial window,
Figure BDA0003475344880000072
is a Boolean matrix (Boolean matrix) in the pre-classification result graph.
S5.4, further traversing all classes to obtain the cumulative probability data of all pixels
Figure BDA0003475344880000073
Can be expressed as a 1 × C vector, then the cumulative probability of all pixels and the sample label data are input into the RF classifier, and F is determined by using the trained modelc(i, j) testing to obtain a classification map of the high spectral data of the grassland community.
In conclusion, the systematic hyperspectral grassland community division method can be used for fully mining the empty-spectral characteristics of hyperspectral data, solving the problems of difficult model parameter adjustment, high training cost and the like of a deep learning method, improving the identification precision of the grassland community on the basis, being completely realized in one step through computer software and having strong practicability.
The invention emphasizes a systematic hyperspectral grassland community dividing method, which can also be considered as a set of complete operation flow aiming at a specific scene, and comprises sample rationality selection, spectral feature extraction, spatial feature extraction, space-spectrum feature fusion and classification post-processing. Based on the above embodiments of the present invention, those skilled in the art should also make modifications and changes (including alternative applications) to the present invention without departing from the principle of the present invention.

Claims (8)

1. A systematic hyperspectral grassland community division method is characterized by comprising the following steps:
s1, selecting partial pixels in the research area range as sample label data;
s2, inputting a hyperspectral image, selecting a proper wave band subset by using a linear prediction LP-based method, and extracting spectral features;
s3, obtaining the spatial characteristics of the hyperspectral image by the selected wave band by using an extended morphology method EMP;
s4, stacking the extracted spectral features and spatial features, fusing in a low-dimensional space by using a Principal Component Analysis (PCA), and obtaining a pre-classification result by combining sample label data and using random forest Radio Frequency (RF);
and step S5, classifying and post-processing the classification result by using a Label Similarity Probability Filter (LSPF) method, and further classifying by using RF (radio frequency) to reduce the influence of salt and pepper noise on the classification result.
2. The systematic hyperspectral grassland community division method according to claim 1, characterized by comprising the following steps: the data in step S1 includes: actual measurement sample point data and vegetation map (1: 1000000) data of the people's republic of China are abbreviated as follows: vegetation map data and national-level natural reserve functional zoning data are abbreviated as follows: protection area data.
3. The systematic hyperspectral grassland community division method according to claim 1 or 2, characterized by: when step S1 is executed, sample label data is selected by actually measuring sample point data, vegetation map data, and protected area data, and combining the image data, and the specific process is as follows:
s1.1, actually measured sample point data is data obtained by manual field investigation, and comprises longitude and latitude coordinates and grassland dominant species information in a 1m x 1m grid, then a ZY1-02D hyperspectral image with the acquisition time similar to that of the actually measured sample point data is selected, a 200m x 200m buffer zone is made according to the actually measured sample point data, and pixels in the buffer zone are used as sample label data;
s1.2, the vegetation map data is published by geological publishing houses in 2008, and the geographical spatial distribution of 11 vegetation type groups, more than 900 groups and sub-groups of 55 vegetation types and more than about 2000 groups dominant species in China is included; the protection area data is a natural centralized distribution area of a country for a representative natural ecosystem and rare or endangered wild animal and plant species, and comprises a core area, a buffer area and an experimental area, wherein the core area is a core area for protecting the ecosystem, is forbidden for any unit and person to enter, and is not allowed to conduct scientific research activities, the buffer area with a certain area is divided at the periphery of the core area, only scientific research observation activities are conducted, and the periphery of the buffer area is divided into the experimental area, so that scientific experiments, teaching practices, visiting investigation, traveling and activities for breeding rare or endangered wild animals and plants can be conducted; overlaying vegetation map data, protection area data and image data in ArcGIS, wherein communities in the core area and the buffer area are considered to be unchanged for many years, but communities in the experimental area need to combine with the current image to select a sample label;
and S1.3, under the precondition of superposition of the vegetation map data, the protected area data and the image data, regarding the special area, mainly comprising the urban area and referring to the vegetation map data, and regarding the community suitable for growing as the community corresponding to the urban area.
4. The systematic hyperspectral grassland community division method according to claim 3, characterized by comprising the following steps: the special area comprises a built-up area, which is a non-agricultural production construction area developed by collecting land and actual construction within the scope of a municipal administration area.
5. The systematic hyperspectral grassland community division method according to claim 1 or 2, characterized by: when step S2 is executed, the method includes performing band selection on the hyperspectral image and extracting spectral features, and includes the specific steps of:
s2.1, selecting a group of wave bands B in the hyperspectral image1And B2Initializing the algorithm and then generating the band subset Φ ═ B1B2}；
Step S2.2, selecting B which is the most different from the current phi from all wave bands by the methods of estimation and linear prediction3And updating the band subset Φ ═ u { B }3}；
a0+a1B1+a2B2＝B′ (1)
Wherein B' is B3Estimation and linear prediction of results of (a)0，a1，a2Is to make the linearity error e | | | B3-the minimum parameter of B' |;
step S2.3, repeating the step S2.2 until phi reaches the preset wave band subset number, and obtaining the spectral characteristic yspe＝Φ。
6. The systematic hyperspectral grassland community division method according to claim 1 or 2, characterized by: when the step S3 is executed, the method includes performing on-off transformation on the hyperspectral image after the band selection, and the specific process includes:
s3.1, performing corrosion-first expansion transformation and then expansion transformation on the hyperspectral pixel x, wherein the operation is defined as follows:
Figure FDA0003475344870000021
wherein, MPγRepresenting the on-transformed image; gamma rayk(x) Representing an open operation performed on image x; l is the number of opening operations;
s3.2, performing expansion-first and corrosion-second closed transformation on the hyperspectral pixel x, wherein the operation is defined as follows:
Figure FDA0003475344870000022
wherein, MPΦRepresenting the closed transformed image; phik(x) Indicating a closing operation performed on the image x; p is the number of closed operations; in equations (2) and (3), k is the size of the structural element, and when k is 0, it means that the original image is not subjected toOperating;
s3.3, combining the original hyperspectral pixel with the result after the opening and closing operation to obtain an EMP result, namely the spatial characteristic ySpa
yspa＝EMP＝{MPγ，x，MPΦ} (4)。
7. The systematic hyperspectral grassland community division method according to claim 1 or 2, characterized by: when step S4 is executed, the method includes stacking spectral features and spatial features of the hyperspectral image, and fusing in a low-dimensional space by using PCA, so as to classify by using RF, which includes the following specific steps:
s4.1, stacking and normalizing the spectral features and the spatial features in the steps S2 and S3 to obtain fused features
Figure FDA0003475344870000031
Step S4.2, data
Figure FDA0003475344870000032
Projection from n-dimensional space to m-dimensional space using PCA:
Figure FDA0003475344870000033
wherein, P ═ P1，...，Pm]M is less than n, and the m characteristic values are selected according to the arrangement sequence of the characteristic values, namely a transformation matrix consisting of characteristic vectors corresponding to the m-dimensional space;
and S4.3, inputting the obtained space-spectrum fusion characteristic G and the sample label data into an RF classifier for training, and testing the image by using the model obtained by training to obtain a grassland community pre-classification result.
8. The systematic hyperspectral grassland community division method according to claim 1 or 2, characterized by: when the step S5 is executed, the obtained pre-classification result is subjected to classification post-processing by using LSPF, and then the classification result of the grassland community is obtained by using the RF classifier, so that salt and pepper noise caused by spectrum similarity is solved;
s5.1, mapping and converting the pre-classification result into a series of binary label graphs according to whether the result pixel belongs to the class c, wherein the value of the result pixel is set to 1 if the result pixel belongs to the class c, and the value of the result pixel is set to 0 if the result pixel does not belong to the class c; if 5 types of ground objects exist in the image scene, the pre-classification result image is divided into 5 layers of binary label images;
step S5.2, within each spatial window, the cumulative probability of pixel (i, j) being of class c is calculated as fc(i, j), then traversing all classes to obtain the cumulative probability data of all pixels as
Figure FDA0003475344870000034
Expressed as a 1 × C vector, then the cumulative probability of all pixels and the sample label data are input into the RF classifier, and the trained model is used to pair FcAnd (i, j) testing to obtain a classification result graph of the high spectral data of the grassland community.
CN202210053671.9A 2022-01-18 2022-01-18 Systematic hyperspectral grassland community division method Active CN114462512B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210053671.9A CN114462512B (en) 2022-01-18 2022-01-18 Systematic hyperspectral grassland community division method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210053671.9A CN114462512B (en) 2022-01-18 2022-01-18 Systematic hyperspectral grassland community division method

Publications (2)

Publication Number Publication Date
CN114462512A true CN114462512A (en) 2022-05-10
CN114462512B CN114462512B (en) 2023-02-10

Family

ID=81408931

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210053671.9A Active CN114462512B (en) 2022-01-18 2022-01-18 Systematic hyperspectral grassland community division method

Country Status (1)

Country Link
CN (1) CN114462512B (en)

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107871132A (en) * 2017-10-31 2018-04-03 广东交通职业技术学院 A kind of hyperspectral image classification method of the adaptive optimizing of space characteristics
CN108614992A (en) * 2018-03-19 2018-10-02 中国地质大学（武汉） A kind of sorting technique of high-spectrum remote sensing, equipment and storage device
CN108846352A (en) * 2018-06-08 2018-11-20 广东电网有限责任公司 A kind of vegetation classification and recognition methods
CN110751142A (en) * 2019-09-25 2020-02-04 河海大学 Improved hyperspectral remote sensing image band selection method
CN112990316A (en) * 2021-03-18 2021-06-18 浪潮云信息技术股份公司 Hyperspectral remote sensing image classification method and system based on multi-saliency feature fusion
CN113762226A (en) * 2021-11-09 2021-12-07 成都理工大学 Method and system for adjusting and improving tree species identification precision based on high spectral resolution

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107871132A (en) * 2017-10-31 2018-04-03 广东交通职业技术学院 A kind of hyperspectral image classification method of the adaptive optimizing of space characteristics
CN108614992A (en) * 2018-03-19 2018-10-02 中国地质大学（武汉） A kind of sorting technique of high-spectrum remote sensing, equipment and storage device
CN108846352A (en) * 2018-06-08 2018-11-20 广东电网有限责任公司 A kind of vegetation classification and recognition methods
CN110751142A (en) * 2019-09-25 2020-02-04 河海大学 Improved hyperspectral remote sensing image band selection method
CN112990316A (en) * 2021-03-18 2021-06-18 浪潮云信息技术股份公司 Hyperspectral remote sensing image classification method and system based on multi-saliency feature fusion
CN113762226A (en) * 2021-11-09 2021-12-07 成都理工大学 Method and system for adjusting and improving tree species identification precision based on high spectral resolution

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
WEIWEI SUN: "A Label Similarity Probability Filter for Hyperspectral Image Postclassification", 《IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING》 *
黄蒂云等: "基于高光谱技术融合图像信息的脱绒棉种品种分类检测研究", 《光谱学与光谱分析》 *

Also Published As

Publication number Publication date
CN114462512B (en) 2023-02-10

Similar Documents

Publication Publication Date Title
Hamylton et al. Evaluating techniques for mapping island vegetation from unmanned aerial vehicle (UAV) images: Pixel classification, visual interpretation and machine learning approaches
Peerbhay et al. Mapping Solanum mauritianum plant invasions using WorldView-2 imagery and unsupervised random forests
Peerbhay et al. Random forests unsupervised classification: The detection and mapping of solanum mauritianum infestations in plantation forestry using hyperspectral data
Mahmon et al. A review on classification of satellite image using Artificial Neural Network (ANN)
CN108197650B (en) Hyperspectral image extreme learning machine clustering method with local similarity maintained
Ünsalan et al. Multispectral satellite image understanding: from land classification to building and road detection
Qader et al. Classification of vegetation type in Iraq using satellite-based phenological parameters
Sasidhar et al. Land cover satellite image classification using ndvi and simplecnn
CN104182767A (en) Active learning and neighborhood information combined hyperspectral image classification method
Tonbul et al. Classification of poplar trees with object-based ensemble learning algorithms using Sentinel-2A imagery
CN114266961A (en) Method for integrating, learning and classifying marsh vegetation stacks by integrating hyperspectral and multiband fully-polarized SAR images
Bektas Balcik et al. Determination of land cover/land use using spot 7 data with supervised classification methods
Bagan et al. Improved subspace classification method for multispectral remote sensing image classification
Du et al. Training SegNet for cropland classification of high resolution remote sensing images
Baidar Rice crop classification and yield estimation using multi-temporal Sentinel-2 data: a case study of terai districts of Nepal
Vignesh et al. A novel multiple unsupervised algorithm for land use/land cover classification
Johnson et al. Gaussianizing the Earth: Multidimensional information measures for Earth data analysis
CN114462512B (en) Systematic hyperspectral grassland community division method
Jijón-Palma et al. Improving stacked-autoencoders with 1D convolutional-nets for hyperspectral image land-cover classification
Jenicka Land Cover Classification of Remotely Sensed Images
Mitrakis et al. A novel self‐organizing neuro‐fuzzy multilayered classifier for land cover classification of a VHR image
Su et al. Discrimination of land use patterns in remote sensing image data using minimum distance algorithm and watershed algorithm
Zhang et al. A rough set decision tree based MLP-CNN for very high resolution remotely sensed image classification
Shi et al. Support Vector Machine for Landscap e Mapping from Remote Sensor Imagery
Dahiya et al. Comparative analysis and implication of Hyperion hyperspectral and landsat-8 multispectral dataset in land classification

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant