Urban impervious layer extraction method based on multi-feature input of hyperspectral remote sensing image
Technical Field
The invention belongs to the technical field of data identification, and particularly relates to a method for extracting an urban impervious bed based on multi-feature input of a hyperspectral remote sensing image.
Background
Currently, the current state of the art commonly used in the industry is such that:
the urban impervious layer (UIS) information is widely applied to the fields of urban land planning, urban 'heat island' and 'carbon island' monitoring, illegal building monitoring, urban garden planning, urban development assessment, urban environment monitoring and analysis and the like. With the continuous deepening of the urbanization process, the UIS information is more and more valued by city planners and government managers, and the extraction method and the application of the UIS also become one of the hotspots in the research field of urban remote sensing. UIS information quality directly affects the reliability of multiple application results.
The hyperspectral image contains hundreds of spectral band information and spatial information, and richer ground spatial features and slight spectral information differences can be obtained by utilizing hyperspectral remote sensing data, so that the characteristic information of different ground objects can be extracted in a discriminative manner. However, the hyperspectral remote sensing data face three problems in surface feature extraction: (1) the spectrum information in the hyperspectral image is easy to be uncertain in computer processing due to the high density of the spectrum, and shadow coverage is easy to be caused on the space structure; (2) the hyperspectral remote sensing image contains less marked sample data, and is not beneficial to supervising and classifying sample selection; (3) more importantly, the hyperspectral remote sensing image contains hundreds of spectral bands, and effective data dimension reduction processing is the premise of hyperspectral image classification processing;
at present, a plurality of dimension reduction methods are used for hyperspectral remote sensing images, and representative algorithms include supervision-based dimension reduction methods, such as Linear Discriminant Analysis (LDA), Local Discriminant Embedding (LDE), Local Fisher Discriminant Analysis (LFDA); unsupervised dimension reduction methods such as Principal Component Analysis (PCA) and the like. In the dimensionality reduction process of the algorithms, some wave band information is abandoned, so that the loss of the wave band information is caused, and the information extraction precision is reduced; moreover, the algorithms can only extract spectral feature information at a lower level, but cannot extract higher-level feature information, such as boundary information, spatial information and the like; for the supervised dimension reduction algorithm, a large amount of marked sample data is needed for training, but the accuracy of the unsupervised dimension reduction algorithm cannot be ensured;
the existing t-SNE algorithm is a nonlinear and unsupervised algorithm, is suitable for data reduction and spectral feature information extraction of high-dimensional remote sensing images, but cannot completely ensure global data optimization and high-precision classification, and has no prediction capability, so the algorithm needs to be further improved;
interference of 'salt and pepper' noise usually occurs in spectral feature extraction, and the phenomenon of wrong separation and missing separation often occurs when ground object classification is carried out by only using the spectral feature extraction, so that the classification precision is reduced;
the existing method for extracting the ground features by mixing the spatial features and the spectral features for supervision and classification is mostly established on the basis of a large number of training samples and testing samples, the classification precision of some algorithms is high, but the influence of artificial sample selection is large, the algorithm has high requirements on the quality of specific hyperspectral remote sensing data, but due to the problem of effective sample selection, the algorithms lack strong universality;
existing advanced Machine learning methods such as Support Vector Machine algorithms (SVMs), Convolutional Neural Networks (CNNs), Deep Belief Networks (DBNs), Convolutional Deep Belief Networks (CDBNs), and the like have high classification accuracy, but have a large amount of redundant data in the learning process, and have high computation time and space complexity, which affects the efficiency of feature extraction. The CDBNs can effectively extract the spatial features and boundary information of the ground features, the requirement on the labeled sample of the input data is not high, and the weight of a connection network between the data input layer and the monitoring layer has certain redundancy.
In summary, the problems of the prior art are as follows:
(1) the hyperspectral remote sensing data face three problems in surface feature extraction: the spectrum information in the hyperspectral image is easy to be uncertain in computer processing due to the high density of the spectrum, and shadow coverage is easy to be caused on the space structure; the hyperspectral remote sensing image contains less marked sample data, and is not beneficial to supervising and classifying sample selection; more importantly, the hyperspectral remote sensing image contains hundreds of spectral bands, and effective data dimension reduction processing is the premise of hyperspectral image classification processing;
(2) in the algorithm dimension reduction process, some band information is abandoned, so that the band information is lost, and the information extraction precision is reduced; moreover, the algorithms can only extract spectral feature information at a lower level, but cannot extract higher-level feature information, such as boundary information, spatial information and the like; for the supervised dimension reduction algorithm, a large amount of marked sample data is needed for training, but the accuracy of the unsupervised dimension reduction algorithm cannot be ensured; for the supervised dimension reduction algorithm, a large amount of marked sample data is needed for training, but the accuracy of the unsupervised dimension reduction algorithm cannot be ensured;
(3) the t-SNE algorithm cannot completely ensure the optimization and high-precision classification of global data, and the algorithm has no prediction capability; the phenomenon of 'salt and pepper' noise is often generated by the spectral feature classification based on the pixels, and the phenomenon of wrong classification and missing classification is often generated by only using the spectral feature extraction to classify the ground objects, so that the classification precision is reduced;
(4) the existing method for extracting the ground objects by mixing the spatial characteristics and the spectral characteristics of supervision and classification is greatly influenced by artificial sample selection, the algorithm has higher requirement on the quality of specific hyperspectral remote sensing data, and the algorithm lacks stronger universality due to the problem of effective sample selection;
(5) the existing advanced machine learning method has a large amount of redundant data in the learning process, has higher complexity of operation time and space, and influences the efficiency of feature extraction; the unsupervised deep learning algorithm CDBNs has low requirements on the labeled samples of the input data, but the weight of the connection network between the data input layer and the monitoring layer has certain redundancy.
The difficulty and significance for solving the technical problems are as follows:
the difficulty in solving the above technical problems lies in the following aspects:
(1) the hyperspectral remote sensing images are rich in spectral information and slight spectral differences, confusion is easy to cause on spectral characteristics, high data dimensionality is achieved, and calculation complexity and classification confusion are obviously improved. The dimension reduction algorithm for effectively reducing spectrum confusion and full-scale low distortion is one of technical difficulties;
(2) the operation time complexity of the existing t-SNE dimension reduction algorithm is high, the prediction capability of hyperspectral remote sensing images with few marked samples is lacked, the time complexity of the algorithm is reduced, and the accuracy of the algorithm for processing supervision and classification of low-marked sample data is improved, which is one of the technical difficulties;
(3) the interference of 'salt and pepper' noise is effectively reduced, the ground feature classification precision is improved, and certain technical difficulty is brought to reducing the data redundancy of the algorithm.
The method solves the technical problems, is favorable for reducing interference and confusion during single spectrum feature extraction, reduces the redundancy of feature extraction data, improves the effect of reducing the dimension of hyperspectral remote sensing image data, and improves the operation efficiency and classification precision of the algorithm.
Disclosure of Invention
Aiming at the problems in the prior art, the invention provides a method for extracting an urban impervious bed based on multi-feature input of a hyperspectral remote sensing image.
The invention discloses a method for extracting an urban impervious layer based on multi-feature input of a hyperspectral remote sensing image, which comprises the following steps of:
the method comprises the steps of performing dimension reduction and spectral feature extraction on hyperspectral remote sensing data based on an Improved t-SNE algorithm (Improved t-SNE), performing spatial feature extraction and boundary information monitoring on a CDBNs algorithm (d-CDBNs) based on a depth compression weight, establishing a ground feature classifier based on multi-feature combined extraction, constructing a UIS extraction model MFCM based on ROSIS-3 hyperspectral remote sensing image multi-feature input, and performing urban impervious layer information extraction.
The invention also aims to provide a method for extracting an urban impervious layer based on multi-feature input of a hyperspectral remote sensing image, which specifically comprises the following steps:
the method comprises the following steps: on the basis of the t-SNE algorithm, an adjacent node searching algorithm is improved, and a hash table searching algorithm is adopted to quickly search out disordered data sets;
step two: the improved algorithm introduces an intra-group correlation coefficient algorithm and carries out similarity evaluation on intra-group nodes; embedding a logistic regression algorithm during the extraction of the spectral features of the algorithm;
step three: the improved t-SNE algorithm reduces the high-dimensional remote sensing data to 2-D data and carries out unsupervised spectral feature extraction;
step four: a depth compression algorithm is introduced to effectively compress the network connection weight between the data input layer and the monitoring layer, and a monitoring CDBNs algorithm is improved;
step five: extracting the spatial features and boundary information of the data after dimensionality reduction through an improved CDBNs algorithm;
step six: and executing a logistic regression classifier by using multiple characteristics such as spectral characteristics, spatial characteristics, boundary information and the like, and extracting the information of the impervious layer in a supervision manner.
Further, the t-SNE algorithm is mainly improved:
(1) a hash table search algorithm is adopted to replace a binary search algorithm;
(2) evaluating the similarity of adjacent nodes, and improving the classification precision;
(3) the embedded logistic regression algorithm increases the predictability of the algorithm and reduces the dependence on the marked sample data.
The invention also aims to provide a computer program for realizing the urban impervious bed extraction method based on the multi-feature input of the hyperspectral remote sensing image.
The invention also aims to provide an information data processing terminal for realizing the urban impervious bed extraction method based on the multi-feature input of the hyperspectral remote sensing image.
Another object of the present invention is to provide a computer-readable storage medium, which includes instructions that, when executed on a computer, cause the computer to execute the method for extracting an impervious bed of a city based on multi-feature input of hyperspectral remote sensing images.
The invention also aims to provide a city impermeable layer extraction control system based on the hyperspectral remote sensing image multi-feature input, which is used for realizing the city impermeable layer extraction method based on the hyperspectral remote sensing image multi-feature input.
In summary, the advantages and positive effects of the invention are:
the advantages and positive effects of the invention are as follows (the result adopts ROSIS-3 hyperspectral data, covers two regions, Pavia unity (U) and Pavia city center (C)):
the time complexity of the hash table searching algorithm is O (1) which is superior to O (lgN) searched by two, so that the algorithm running time is saved on the searching of adjacent nodes, and the algorithm running efficiency is improved;
in order to improve the similarity of adjacent nodes and further improve the classification precision, the improved algorithm introduces an intra-group correlation coefficient (ICC) algorithm, carries out similarity evaluation on the intra-group nodes, improves the similarity in the adjacent node groups, and improves the distinguishing degree of different surface feature characteristics due to approximate spectral reflection, thereby improving the precision of spectral feature extraction;
according to the invention, a Logistic Regression (LR) algorithm is embedded during the extraction of the spectral features of the algorithm, so that on one hand, the classification precision in the feature extraction process is increased, on the other hand, the defect that the original algorithm lacks predictability is overcome, and the dependence on marking sample data is reduced in the unsupervised classification process;
the improved t-SNE algorithm is mainly responsible for dimensionality reduction and spectral feature extraction of the hyperspectral remote sensing image. Reducing the high-dimensional remote sensing data to 2-D data, and performing unsupervised spectral feature extraction;
the invention introduces a depth compression algorithm to effectively compress the network connection weight between the data input layer and the monitoring layer, thereby reducing the redundancy of the weight, establishing the optimal sharing weight, reducing the complexity of the operation time and space and improving the performance of the algorithm; the spatial and temporal complexity of the detection of spatial features and boundary information is improved;
the improved CDBNs algorithm is responsible for extracting the spatial characteristics of the data after dimension reduction and monitoring the boundary information;
the invention provides a UIS extraction model (MFCM) for multi-feature mixed classification of spectral features, spatial features, boundary information and the like; the model extracts the information of the impervious layer from the multi-feature comprehensive information through a supervised logistic regression classifier.
The invention improves t-SNE and CDBNs algorithms, and constructs a spectral feature, spatial feature and boundary information combined UIS extraction model through a logistic regression classifier, thereby obtaining better classification precision; the model effectively utilizes the spectral characteristics, the spatial characteristics, the boundary information and other comprehensive characteristics for extraction, and effectively reduces the influence of salt and pepper noise.
Drawings
FIG. 1 is a flow chart of a method for extracting an urban impervious layer based on multi-feature input of a hyperspectral remote sensing image according to an embodiment of the invention.
FIG. 2 is a UIS extraction model framework diagram based on ROSIS-3 hyperspectral remote sensing image multi-feature input provided by the embodiment of the invention.
FIG. 3 is a schematic diagram of classification results of the Pavia unity data under different dimensions comparing the improved t-SNE algorithm provided by the embodiment of the present invention with the commonly used dimension reduction algorithms LDA, LFDA, LDE, PCA, and t-SNE.
FIG. 4 is a diagram illustrating classification results of the Pavia city center data in different dimensions by comparing the improved t-SNE algorithm provided by the embodiment of the present invention with the commonly used dimension reduction algorithms LDA, LFDA, LDE, PCA, and t-SNE.
FIG. 5 is a schematic diagram of classification accuracy under different dimensions and different puzzles of the improved t-SNE algorithm provided by the embodiment of the invention.
FIG. 6 is a schematic diagram of the running time of the improved t-SNE algorithm and the original t-SNE algorithm for two data sets under different confusion conditions provided by the embodiment of the invention.
FIG. 7 is a diagram illustrating the results of the extraction of the impermeable layers for two data sets using several algorithms according to an embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail with reference to the following embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
The invention improves t-SNE and CDBNs algorithms, and constructs a spectral feature, spatial feature and boundary information combined UIS extraction model through a logistic regression classifier, thereby obtaining better classification precision.
The invention provides a UIS extraction model based on multi-feature information input of a ROSIS-3 hyperspectral remote sensing image. The model adopts an improved t-distributed random neighborhood embedding algorithm (t-SNE), an improved convolution depth trust network algorithm (CDBNs) and a logistic regression model (LR), and utilizes the spectral, spatial and edge characteristics extracted from the hyperspectral remote sensing image to construct a UIS information classification extraction strategy.
The urban impervious layer extraction method based on the multi-feature input of the hyperspectral remote sensing image comprises a UIS extraction model based on the multi-feature input of the ROSIS-3 hyperspectral remote sensing image, wherein the model comprises dimension reduction and spectral feature extraction of hyperspectral remote sensing data based on an improved t-SNE algorithm; extracting spatial features and monitoring boundary information of a CDBNs algorithm (d-CDBNs) based on depth compression weights; and (4) a ground object classifier with multi-feature joint extraction.
The application principle of the present invention will be described in detail with reference to the accompanying drawings;
as shown in fig. 1, the method for extracting an urban impervious layer based on multi-feature input of a hyperspectral remote sensing image, provided by the embodiment of the invention, specifically comprises the following steps:
s101: on the basis of the t-SNE algorithm, an adjacent node searching algorithm is improved, and a hash table searching algorithm is adopted to quickly search out disordered data sets;
the time complexity of the existing hash table searching algorithm is O (1) which is superior to O (lgN) searched by two, so that the algorithm running time is saved on the searching of adjacent nodes, and the algorithm running efficiency is improved;
s102: the improved algorithm introduces an intra-group correlation coefficient algorithm and carries out similarity evaluation on intra-group nodes;
embedding a logistic regression algorithm during the extraction of the spectral features of the algorithm; the intra-group correlation coefficient algorithm improves the similarity in adjacent node groups and improves the distinguishing degree of different surface feature characteristics similar to spectral reflection, thereby improving the accuracy of spectral feature extraction;
s103: the improved t-SNE algorithm reduces the high-dimensional remote sensing data to 2-D data and carries out unsupervised spectral feature extraction;
the improved algorithm increases a prediction function, on one hand, the classification precision in the feature extraction process is increased, on the other hand, the defect that the original algorithm is lack of predictability is overcome, and the dependence on marking sample data is reduced in the unsupervised classification process;
s104: a depth compression algorithm is introduced to effectively compress the network connection weight between the data input layer and the monitoring layer, and an unsupervised CDBNs algorithm is improved;
the invention introduces a depth compression algorithm to effectively compress the network connection weight between the data input layer and the monitoring layer, thereby reducing the redundancy of the weight, establishing the optimal sharing weight, reducing the complexity of the operation time and space and improving the performance of the algorithm;
s105: performing spatial feature extraction and boundary information monitoring on the data after dimension reduction through an improved CDBNs algorithm;
the improved CDBNs algorithm extracts the spatial characteristics and boundary information of the 2-D data set through convolution operation, maximum pooling operation and sparse rules;
s106: extracting watertight layer information from multi-feature comprehensive information through a supervised logistic regression classifier by using a UIS extraction model (MFCM) for carrying out mixed classification on multi-features such as spectral features, spatial features, boundary information and the like;
the MFCM is an abbreviation of a Multi-features Cooperation Model, the Model is used for extracting a hyperspectral remote sensing image impervious layer by combining spatial features, boundary information and spectral features, and the Multi-feature extraction method is helpful for reducing confusion and interference of 'salt and pepper' noise in the spectral feature extraction process.
The t-SNE algorithm provided by the embodiment of the invention is mainly improved as follows:
(1) a hash table search algorithm is adopted to replace a binary search algorithm;
(2) evaluating the similarity of adjacent nodes, and improving the classification precision;
(3) the embedded logistic regression algorithm increases the predictability of the algorithm and reduces the dependence on the marked sample data.
As shown in fig. 2, the UIS extraction model framework diagram based on ROSIS-3 hyperspectral remote sensing image multi-feature input provided by the embodiment of the invention.
The invention provides a UIS extraction model (MFCM) for multi-feature mixed classification of spectral features, spatial features, boundary information and the like; the model extracts the information of the impervious layer from the multi-feature comprehensive information through a supervised logistic regression classifier.
The application principle of the present invention is further explained with reference to the following specific embodiments;
example 1;
the main technology of the invention comprises the following seven aspects:
(1) on the basis of the t-SNE algorithm, the invention improves the adjacent node searching algorithm and adopts the hash table searching algorithm to replace the original binary search algorithm. The Hash table algorithm can quickly search out an unordered data set, the binary search is a search algorithm established on the basis of an ordered data set, and similar data points of the hyperspectral remote sensing image data set are unordered. The time complexity of the hash table searching algorithm is O (1) which is superior to O (lgN) searched by two, so that the algorithm running time is saved on the searching of adjacent nodes, and the algorithm running efficiency is improved;
(2) in order to improve the similarity of adjacent nodes and further improve the classification precision, an intra-group correlation coefficient (ICC) algorithm is introduced into the improved algorithm, the similarity evaluation is carried out on the intra-group nodes, the similarity in the adjacent node groups is improved, the difference of different surface feature characteristics similar to spectral reflection is improved, and the precision of spectral feature extraction is improved;
(3) in order to improve the predictability of the t-SNE algorithm and reduce the dependence on marked sample data, a Logistic Regression algorithm (LR) is embedded during the extraction of the spectral features of the algorithm, so that on one hand, the classification precision in the feature extraction process is improved, on the other hand, the defect that the original algorithm lacks predictability is overcome, and the dependence on the marked sample data is reduced in the unsupervised classification process;
(4) the improved t-SNE algorithm is mainly responsible for dimensionality reduction and spectral feature extraction of the hyperspectral remote sensing image. Reducing the high-dimensional remote sensing data to 2-D data, and performing unsupervised spectral feature extraction;
(5) the invention improves the unsupervised CDBNs algorithm. In order to reduce the weight redundancy among the visible layer, the hidden layer and the pooling layer of the CDBNs algorithm. The invention introduces a depth compression algorithm to effectively compress the network connection weight between the data input layer and the monitoring layer, thereby reducing the redundancy of the weight, establishing the optimal sharing weight, reducing the complexity of the operation time and space and improving the performance of the algorithm;
(6) the improved CDBNs algorithm is responsible for extracting the spatial features of the data after dimensionality reduction and monitoring boundary information;
(7) the invention provides a UIS extraction model (MFCM) for multi-feature mixed classification of spectral features, spatial features, boundary information and the like. The model extracts the information of the impervious layer from the multi-feature comprehensive information through a supervised logistic regression classifier.
The application results of the present invention will be further explained below with reference to specific experiments;
experiment 1;
the invention constructs a ROSIS-3 hyperspectral remote sensing image multi-feature input-based UIS extraction model (MFCM), which mainly comprises three parts: (1) the hyperspectral remote sensing data dimensionality reduction and spectral feature extraction based on an Improved t-SNE algorithm (Improved t-SNE); (2) extracting spatial features and monitoring boundary information of a CDBNs algorithm (d-CDBNs) based on depth compression weights; (3) and (4) a ground object classifier with multi-feature joint extraction.
The invention utilizes ROSIS-3 hyperspectral remote sensing data sets (Pavia unity and Pavia city center) to carry out experiments and algorithm performance tests, and adopts the main performance indexes of a computer as follows: a CPU: inter core i5-4200U, dominant frequency: 2.3GHz, 4 cores, 4 threads, 8GB RAM, Windows 7 operating system (64 bits). The performance evaluation indexes include production Precision (PA), user precision (UA), Kappa coefficient (Kappa), overall precision (OA), average precision (AA), and the like. The main effects are as follows:
(1) the highest overall classification precision of the improved t-SNE algorithm obtained by an LR classifier is 82.63%, which is 4.22% higher than that of the original t-SNE algorithm, and is 7.13% higher than that of other commonly used dimension reduction algorithms (LDA, LDE, LFDA, PCA, t-SNE) on average, and the classification results of the algorithms are shown in Table 1 and FIGS. 3 and 4. Table 1 shows the results of the LR classifier, the improved t-SNE algorithm and the conventional dimensionality reduction algorithm when the dimensionality reduction of the data set is 2-D
As shown in fig. 3, the improved t-SNE algorithm provided by the embodiment of the present invention is compared with the commonly used dimension reduction algorithms LDA, LFDA, LDE, PCA, and t-SNE, and a schematic diagram of classification results of Pavia unity data under different dimensions is shown.
As shown in FIG. 4, the improved t-SNE algorithm provided by the embodiment of the invention is compared with the commonly used dimensionality reduction algorithms LDA, LFDA, LDE, PCA and t-SNE, and a schematic diagram of classification results of the Pavia city center data under different dimensions is shown.
And (3) comparing the classification results of the two data sets by adopting an LR classifier and an improved t-SNE algorithm with a common algorithm under different dimensions. The classification results obtained by the LR classifier when the dimensions of the two data sets are reduced to 2-D are shown in Table 1: the improved t-SNE algorithm is compared to the commonly used dimension reduction algorithms (LDA, LDE, LFDA, PCA, t-SNE).
(2) The classification precision of the improved t-SNE algorithm under the conditions of different dimensions and different puzzles is shown in FIG. 5, and the dimension range is tested: 2-12, confusion range: 10-100, the two data sets obtain the highest test precision when the confusion degree is 50 and the dimensionality is 6;
as shown in FIG. 5, the classification accuracy of the improved t-SNE algorithm is schematically shown under the conditions of different dimensions and different puzzles.
In the range of the test data, when the confusion degree is 50 and the dimensionality is 6, the test data obtains the highest classification precision, (a) the Pavia unity data, wherein the highest classification precision is 84.23%; (b) the highest classification accuracy of the Pavia city center data is 87.87%.
(3) In the time complex test, a t-SNE algorithm and an improved t-SNE algorithm are tested, the dimensionality of two data sets is reduced to 2-D respectively, the iteration times are 1000 times, the running time of the two algorithms under different puzzles is shown in figure 6, and the time complexity of the improved t-SNE algorithm is averagely reduced by 17.23 seconds compared with that of the original t-SNE algorithm;
as shown in FIG. 6, the improved t-SNE algorithm and the original t-SNE algorithm run time diagrams for two data sets under different confusion conditions provided by the embodiment of the invention.
U:Pavia university;C:Pavia city center。
(4) The time complexity and the space complexity of a CDBNs algorithm (d-CDBNs) for deeply compressing the data input layer and monitoring the interlayer network connection weight are reduced, the maximum time of the d-CDBNs algorithm is respectively reduced to 74.9 seconds and 27.4M in the running time and the space relative to the original CDBNs algorithm, and the classification precision of the algorithm is not influenced after the weight is compressed.
The compression is 4 times, the maximum pooling layer constant C is 2, two data sets are respectively tested, and the results are shown in tables 2 and 3:
table 2: CDBNs algorithm and modified algorithm (d-CDBNs) tested the results in the Pavia unity scene.
Table 3: the CDBNs algorithm and the modified algorithm (d-CDBNs) tested the results at the Pavia city center.
(5) The UIS extraction model (MFCM) provided by the invention mixes spectral characteristics, spatial characteristics and boundary characteristics, and forms a multi-characteristic combined impervious layer extraction strategy by utilizing an LR classifier. The model respectively carries out the extraction test of the impervious layer on the two data sets, and the overall classification precision of the model is 16.11 percent higher than the extraction precision of single spectral feature on average. To demonstrate the performance of the MFCM model, the MFCM model was experimentally tested and compared to commonly used more advanced machine learning algorithms (SVM, CNNs, DBNs and CDBNs). The experimental result shows that the precision of the MFCM model is improved by not less than 4.6 percent compared with the precision of the conventional algorithm. The results of these algorithms on the two dataset imperviousness layer extraction are shown in fig. 7.
As shown in fig. 7, the embodiment of the present invention provides a schematic diagram of the result of extracting the impermeable layer of two data sets by using several algorithms.
In the figure, (a) and (g) are two data false color images, respectively, band (46, 27, 10) and band (85, 60, 5), (b) - (f) are Pavia unity; (h) - (l) is a Pavia city center, in the following order: SVM, CNNs, DBNs, CDBNs algorithms and MFCM models.
The test data of the invention adopts two kinds of hyperspectral remote sensing data of a German ROSIS-3 sensor, the space resolution of two images is 1.3m, the electromagnetic spectrum range is from 0.43 to 0.86 μm, wherein, the Pavia unity data set selects 103 effective wave bands, each wave band has 610 × 340 pixels, the Pavia city center data set selects 102 effective wave bands, and each wave band has 1096 × 715 pixels.
The invention has different accuracy due to different algorithm initialization parameters, and the main parameters related in the algorithm are as follows: the method comprises the following steps of adjacent node confusion, iteration times, dimensionality reduction, correlation coefficient, hidden layer number, depth compression ratio, pixel filtering window size and the like.
In the above embodiments, the implementation may be wholly or partially realized by software, hardware, firmware, or any combination thereof. When used in whole or in part, can be implemented in a computer program product that includes one or more computer instructions. When loaded or executed on a computer, cause the flow or functions according to embodiments of the invention to occur, in whole or in part. The computer may be a general purpose computer, a special purpose computer, a network of computers, or other programmable device. The computer instructions may be stored in a computer readable storage medium or transmitted from one computer readable storage medium to another, for example, the computer instructions may be transmitted from one website site, computer, server, or data center to another website site, computer, server, or data center via wire (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL), or wireless (e.g., infrared, wireless, microwave, etc.)). The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device, such as a server, a data center, etc., that includes one or more of the available media. The usable medium may be a magnetic medium (e.g., floppy Disk, hard Disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., Solid State Disk (SSD)), among others.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents and improvements made within the spirit and principle of the present invention are intended to be included within the scope of the present invention.