CN101034439A - Remote sensing image classify method combined case-based reasoning with Fuzzy ARTMAP network - Google Patents

Abstract

The invention is a kind of remote sensing image classification method combinated case-based reasoning with and Fuzzy ARTMAP network, belongs to recognition field of a computer remote sensing image processing and pattern. Its characteristic is that the following steps: First establishes reflecting the new paradigm of unclassified remote sensing image characteristics, searches example index set that is suitable to the current classification problem by using two-tier strategy of characteristics retrieval in space-time and selected contents features from examples warehouse; Secondly constructs classification program according to paradigm index set and makes further adjustments and amendments points, unclassified remote sensing image is operated in classification by the classification program after amendment;finally the classification program after amendment will be amended into the classification program for example by the use of updated strategy of paradigm index set, to prepare for classified reference of the next operation. The effects and benefits of this invention is:it can fully tap and optimize in conformity with the integration of remote sensing samples and knowledge of classification, substantially improve classification accuracy of remote sensing image supervision under less training samples.

Description

The classification of remote-sensing images method that a kind of case-based reasoning combines with Fuzzy ARTMAP network

Technical field

The present invention relates to the computing machine remote sensing image and handle and area of pattern recognition, is the classification of remote-sensing images method that a kind of case-based reasoning combines with Fuzzy ARTMAP network.The present invention can reuse and optimize and combine existing remote sensing sample data and classificating knowledge, being particularly suitable for sampling on the spot has a supervision classification of remote-sensing images problem under the ageing situation such as relatively poor of difficulty, training sample, all can be widely used in fields such as Remote Sensing Information Processing System, the automatic deciphers of remote sensing image.

Background technology

Supervised classification method is the main method that classification of remote-sensing images work is adopted, and most supervised classification method is (Stathakis, the D. ， ﹠amp that is based upon under the sufficient reliable premise hypothesis of training sample; Vasilakos, A. (2006) .Comparison ofcomputational intelligence based classification techniques for remotely sensed optical imageclassification.IEEE Transactions on Geoscience and Remote Sensing, 44 (8), 2305-2318).But in actual applications, the sample collection need of work of remote sensing image consumes great amount of manpower and material resources and financial resources, and time span is bigger, especially for regional all the more so (Baraldi, A., the Bruzzone of people such as marsh, wetland for the on-the-spot investigation difficulty, L. ， ﹠amp; Blonda, P. (2005) .Quality assessment of classification and cluster maps without ground truth knowledge.IEEE Transactions on Geoscience and Remote Sensing, 43 (4), 857-872).On the one hand, the limited amount and the renewal frequency of remote sensing sample are lower, are difficult to adapt to the needs that supervised classification is arranged, and have seriously influenced the order of accuarcy of classification.On the other hand, lack effective means and existing sampled data and classificating knowledge are rationally reused and optimize, caused significant wastage (Foody, G.M., Mathur, A., Sanchez-Hernandez, the C. ， ﹠amp of resource; Boyd, D.S. (2006) .Training set sizerequirements for the classification of a specific class.Remote Sensing of Environment, 104 (1), 1-14).Though do not have the supervised classification method problem of avoidance training sample to a certain extent, it is excessive to the degree of dependence of initial cluster center not have supervised classification method, the classifying quality instability.

At present, mostly be statistical method at the solution of the problems referred to above, as Cascade sorting technique (Bruzzone, L. ， ﹠amp based on greatest hope (EM); Cossu, R. (2002) .A multiple-cascade-classifier system for a robust andpartially unsupervised updating of land-cover maps.IEEE Transactions on Geoscience and RemoteSensing, 40 (9), 1984-1996), the principle of work of these methods is: based on a classified remote sensing image information, in conjunction with the unknown categorical data of unfiled remote sensing image, ask for the statistical sorter parameter that satisfies the greatest hope condition simultaneously; Utilize these parameters of trying to achieve to make up sorter, unfiled image is carried out sort operation.There is following deficiency in these methods: (1) requires data to satisfy the normal distribution prerequisite, and the distribution situation complexity of actual remotely-sensed data, is difficult to guarantee normal state character; (2) require the classification number known and immobilize, and often occur the different spectrum phenomenon of jljl (subclass that promptly has feature difference with kind inside) in the practical application, the number of subclass is difficult to determine; (3) be primarily aimed at operation between two width of cloth images, its operand is big and lack the available strategy that solves several images.

Summary of the invention

The objective of the invention is to solve under the less situation of training sample remote sensing image supervised classification problem and to the integrated optimization problem of existing remote sensing sample, classificating knowledge.The classification of remote-sensing images method that a kind of case-based reasoning combines with Fuzzy ARTMAP network is proposed.

Technical scheme of the present invention is as follows

This method comprises four major parts: example structure, case retrieval, case reuse and correction, case library upgrade.Concrete steps of the present invention are as follows:

1. make up new example: with the space-time characteristic (coverage and acquisition time) of unfiled remote sensing image space-time characteristic as new example; The sample data set (still not having classification information) of unfiled remote sensing image is gathered as the sample data of new example.Obtain reflecting the new example of unfiled remote sensing image feature through above-mentioned two steps operation.

2. carry out the space-time characteristic retrieval: the space-time characteristic with new example is a foundation, utilize the space-time similarity function to carry out the space-time characteristic search operaqtion, retrieve in the case library and all satisfying the example set of setting threshold requirement aspect coverage and the acquisition time two with new example from the coverage coincidence degree of remote sensing image, degree of approximation two aspects of remote sensing image acquisition time.

3. it is selected to carry out content characteristic: utilize the classification model in the example that preliminary search goes out that new example sample data is carried out sort operation.The corresponding classification results that obtains with each example is a foundation, and the center weight vector of utilization calculates the retrieval energy function of each classification of each example correspondence.Each classification is asked for the example index value with minimum retrieval energy function, set up paradigm index set after selected with this.

4. be fundamental construction preliminary classification scheme with the paradigm index set after selected: the classification model (comprising Fuzzy ARTMAP network weight and corresponding with it mode top weights) of respective classes in the example of index value correspondence is extracted, be reassembled as the preliminary classification scheme.

5. be foundation with the data characteristics in the new example, revise the preliminary classification scheme:

(1), can guarantee that like this each cluster centre can both corresponding definite classification logotype with the center weights of (containing subclass) of all categories in the classification model preliminary examination cluster centre as Fuzzy c-means sorter.

(2) according to Fuzzy c-means iterative algorithm the data in the new example are carried out cluster operation, obtain (containing subclass) of all categories cluster centre after the algorithm convergence, also obtain the preliminary classification result of new example data simultaneously according to new example data feature adjustment.

(3) cluster centre of the Fuzzy c-means after will restraining is as the mode top weights of classification schemes; The network weight that utilizes the preliminary classification result of new example data to train Fuzzy ARTMAP in the classification schemes again.So far obtain revised classification of remote-sensing images scheme.

6. utilize the Fuzzy ARTMAP network weight in the classification of remote-sensing images scheme that remote sensing image is carried out sort operation, sorting algorithm is according to the sorting algorithm of Fuzzy ARTMAP network.Obtain the classification of remote-sensing images result.

7. with the space-time characteristic of new example (obtaining by (1) step), revised classification of remote-sensing images scheme (obtaining by (5) step) and sorted remote sensing image data (by the acquisition of (6) step) chief component as the example that will store,

This example is deposited in the case library in order to sort operation reference next time, finish the renewal operation of case library.

The invention has the beneficial effects as follows case-based reasoning is combined with Fuzzy ARTMAP network, realize the mutual supplement with each other's advantages of two kinds of methods.Case-based reasoning for existing remotely-sensed data and classificating knowledge reuse, integrate and optimization provides the high-efficiency method guidance, both can handle the knowledge utilization between two width of cloth remote sensing images, the remote sensing image that also is particularly suitable for the multidate big data quantity is handled operation; Fuzzy ARTMAP network has characteristics such as subclass dynamically generates, the clear correspondence of weights classification, can strengthen the Knowledge Extraction and the inducing ability of case-based reasoning method.Sorting technique that the present invention carries is not to the specific (special) requirements of sample distribution situation, can fully reuse and optimize existing sample data and classificating knowledge, pay attention to simultaneously merging mutually with the information of unfiled remote sensing image, generate suitable classification schemes, can solve the remote sensing image supervised classification problem under the less situation of training sample preferably and improve nicety of grading significantly.

Description of drawings

Fig. 1 is an operational flow diagram of the present invention.

Among Fig. 1, example makes up the file layout that part is used for design sample data and classification model (knowledge), sets up the new example of the unfiled image of reflection simultaneously; The case retrieval part can retrieve the example set of the most suitable current image classification problem from case library, and therefrom extracts the paradigm index set that makes up classification schemes; Case reuse and makeover process can be set up classification schemes at current image according to paradigm index set, simultaneously classification schemes are estimated and revised the classificating requirement that makes it to meet current image; Case library upgrades classification schemes is stored in case library with the form of example, has realized the learning functionality to new image information.

Fig. 2---Fig. 5 is multispectral remote sensing image and a classification results comparison diagram thereof in the embodiment of the present invention.

Fig. 2 is a unfiled multispectral remote sensing image in the embodiment of the present invention, and acquisition time is on August 28th, 2003, and coverage is 46 ° of 09 ' N～47 ° 22 ' N; 123 ° of 24 ' E～124 ° 29 ' E; Fig. 3 is the classification results of SVM method; Fig. 4 is the classification results of Cascade method; The classification results of Fig. 5 for adopting the present invention to obtain.

Embodiment

In order to understand technical scheme of the present invention better, embodiments of the present invention are further described below in conjunction with accompanying drawing.

The present invention is an example with multispectral remote sensing image shown in Figure 2, according to method flow shown in Figure 1 it is carried out sort operation,

Concrete implementation step is as follows:

1. example building process

The all corresponding width of cloth remote sensing image of each example in the case library, its normal structure constitutes (seeing formula (1)) by three parts: space-time characteristic, data acquisition and classification model set.

C＝(Φ，D，P) (1)

Wherein, C represents example, and Φ represents space-time characteristic, and D represents the sample data set, the set of P presentation class template.Each several part all comprises concrete content in the example, shown in formula (2)-(4):

Φ＝(LA，T) (2)

Wherein, LA represents the coverage of the corresponding image of example, and T represents the acquisition time of corresponding image.

D＝{x _i}；i∈[1，N] (3)

Wherein, x _iSample in the expression data acquisition, N represents the total sample number in the data acquisition.

P＝(W，V)＝{(W _j，V _j)|W _jW，V _jV；j∈[1，L]}

＝{(w _jh，v _jh)|w _jh∈W _j，v _jh∈V _j；h∈π _j，j∈[1，L]} (4)

Wherein, W represents the weight vector of Fuzzy ARTMAP network, V represent to add with W center weight vector one to one.π _jExpression belongs to the subclass set of j class, and L represents classification sum, w _JhExpression belongs to the FuzzyARTMAP network weight of h subclass in the j class, v _JhExpression and w _JhCorresponding center weights.

But the present invention utilizes the weights interpretation characteristic (weights of each node correspondence are represented specific classification or subclass feature) of Fuzzy ARTMAP network, with the weight vector of Fuzzy ARTMAP network as classification model.Because Fuzzy ARTMAP network adopts fuzzy intersection operation symbol in training process, the weight vector monotone decreasing does not reflect moving of mode top, so the weights of Fuzzy ARTMAP network can not directly reflect the center of input pattern.The present invention has introduced new mode top weight vector V (referring to formula (4)), adjusts the center weights in the FuzzyARTMAP training:

$v_{\mathrm{jh}}=v_{\mathrm{jh}}+\frac{1}{N_h}(x-v_{\mathrm{jh}});h\∈{\mathrm{\π}}_j,j\∈[1,L]---\left(5\right)$

Wherein, N _hExpression is divided into the number of samples of j class h subclass, L, v _JhAnd π _jDefinition referring to formula (4).

Make up the example of known image correspondence according to formula (1)-(4); For the unfiled image of new example correspondence, the sample data that only needs to extract its no classification information gets final product, and the classification model of new example can be generated automatically by the step of back.

2. case retrieval process

The present invention adopts double-deck search strategy, and promptly space-time characteristic key and content characteristic are selected.Generate through the case retrieval process and to mate paradigm index set most, for the foundation of classification schemes provides foundation.Table 1 is depicted as the space-time characteristic information of existing example in the case library.

1. carry out the space-time characteristic retrieval of example according to formula (6)

$\mathrm{\Ξ}=\left\{C^k\right|\frac{S\left({\mathrm{LA}}^{\mathrm{new}}I{\mathrm{LA}}^k\right)}{S\left({\mathrm{LA}}^{\mathrm{new}}\right)}\≥B_{\mathrm{LA}},|T^{\mathrm{new}}-T^k|\≤B_T;k\∈[1,k]\}---\left(6\right)$

Wherein, Ξ represents in the case library and the enough similar example set (containing underscore in the table 1 partly is satisfactory example) on spatial coverage and time of new example, LA ^NewThe coverage of representing the unfiled image of new example correspondence, LA ^kThe coverage of k the corresponding image of example in the expression case library, S () expression area coverage computing function, LA ^New∩ LA ^kThe intersection of representing unfiled image and the corresponding image coverage of example k, B _LAExpression coverage similarity degree threshold value, T ^NewThe acquisition time of representing unfiled image, T ^kThe acquisition time of the corresponding image of expression example k, B _TExpression image acquisition time similarity threshold value.For ease of further expressing and deriving, define the index set Γ of example collection Ξ with formula (7):

Γ＝{τ|C ^τ∈Ξ} (7)

Wherein, Γ represents the index set of example collection Ξ, and τ represents to gather example C among the Ξ ^τIndex value.

The space-time characteristic information of existing example in table 1 case library

Index value in the case library	The acquisition time of corresponding remote sensing image	Coverage
			1 2 3 4 5 6 7 8	1986/11/05 1989/08/09 1992/05/14 1996/04/28 1998/05/04 1999/10/08 2000/09/24 2001/09/27	46°09′N～47°22′N；123°24′E～124°29′E 46°44′N～47°39′N；123°32′E～124°50′E 46°30′N～47°19′N；123°27′E～124°22′E 46°44′N～47°12′N；123°30′E～124°11′E 46°46′N～47°38′N；123°31′E～124°52′E 46°44′N～47°39′N；123°32′E～124°51′E 46°44′N～47°39′N；123°32′E～124°49′E 46°43′N～47°38′N；123°31′E～124°50′E

Underscore is partly for retrieving the satisfactory example in back in the table through space-time characteristic

2. utilize each example C among the example set Ξ ^τClassification model the unfiled data in the new example are classified, sorting algorithm adopts the sorting algorithm of Fuzzy ARTMAP network, obtains by C ^τThe interim classification results that provides:

$D^{\mathrm{\τ}}=\{D_f^{\mathrm{\τ}};\mathrm{\τ}\∈\mathrm{\Γ},j\∈[1L\left]\right\}=\left\{D_{\mathrm{jh}}^{\mathrm{\τ}}\right|D_{\mathrm{jh}}^{\mathrm{\τ}}\⊆D_j^{\mathrm{\τ}};\mathrm{\τ}\∈\mathrm{\Γ},j\∈\left[1L\right],h\∈{\mathrm{\π}}_j^{\mathrm{\τ}}\}---\left(8\right)$

Wherein, D ^τThe interim classification results set that expression is provided by example τ, D ^j _τExpression D ^τIn belong to the interim classification results set of j class, π ^j _τExpression D ^τIn belong to the subclass set of classification j, D _Jh ^τExpression belongs to the interim classification results set of subclass h among the classification j.

3. the interim classification results that obtained according to the last step, calculate the retrieval energy function of each example among the Ξ:

$E_j^{\mathrm{\τ}}=\frac{1}{N_j^{\mathrm{\τ}}}\underset{h\∈{\mathrm{\π}}_j^{\mathrm{\τ}}}{\mathrm{\Σ}}\underset{x\∈D_{\mathrm{jh}}^{\mathrm{\τ}}}{\mathrm{\Σ}}{\left|\right|x-v_{\mathrm{jh}}^{\mathrm{\τ}}\left|\right|}^2;\mathrm{\τ}\∈\mathrm{\Γ},j\∈[1,L]---\left(9\right)$

Wherein, E _j ^τThe retrieval energy function of classification j among the expression example τ, N _j ^τRepresent to belong in the interim classification results of corresponding example τ the number of samples of classification j, π _j ^τAnd D _Jh ^τDefinition referring to formula (8), v _Jh ^τThe center weights of h subclass in the j class of expression example τ, L and Γ are respectively referring to formula (4) and formula (7).

4. the result who obtains based on above-mentioned steps carries out the selected operation of content characteristic, generates to mate paradigm index set in the case library most:

$\mathrm{\Ψ}=\left\{{\mathrm{\ψ}}_j\right|{\mathrm{\ψ}}_j=\underset{\mathrm{\τ}\∈\mathrm{\Γ}}{\arg \min }\left\{E_j^{\mathrm{\τ}}\right\};j\∈[1,L]\}---\left(10\right)$

Wherein, Ψ represents the coupling paradigm index set in the case library, ψ _jThe example index value of j class classification model correspondence in the presentation class scheme.Γ is referring to formula (7), E _j ^τReferring to formula (9), L is referring to formula (4).Table 2 is depicted as in the practical operation the coupling paradigm index set that obtains through above-mentioned case retrieval process.

The coupling paradigm index set that table 2 obtains after the selected double-deck search strategy of space-time characteristic retrieval and content characteristic

The atural object classification	The example index value
		Nonirrigated farmland, reed grass glade, water body alkaline land	8 2 8 6 7 8

3. case reuse and makeover process

According to mating most paradigm index set Ψ, make up classification schemes according to formula (11)-(13):

$P^S=(W^S,V^S)=\{(W_j^S,V_j^S);j\∈[1,L\left]\right\}$

$=\left\{(w_{\mathrm{jh}}^S,v_{\mathrm{jh}}^S)\right|w_{\mathrm{jh}}^S\∈W_J^S,v_{\mathrm{jh}}^S\∈V_j^S;h\∈{\mathrm{\π}}_j^S,j\∈[1,L]\}$ (11)

$w_j^S=w_{\mathrm{jh}}^{{\mathrm{\ψ}}_j},v_{\mathrm{jh}}^S=v_{\mathrm{jh}}^{{\mathrm{\ψ}}_j};h\∈{\mathrm{\π}}_j^{{\mathrm{\ψ}}_j},{\mathrm{\ψ}}_j\∈\mathrm{\Ψ},j\∈[1,L]---\left(12\right)$

$P^S=\left\{(w_{\}\mathrm{jh}}^{{\mathrm{\ψ}}_j},v_{\mathrm{jh}}^{{\mathrm{\ψ}}_j})\right|w_{\mathrm{jh}}^{{\mathrm{\ψ}}_j}\∈W_J^{{\mathrm{\ψ}}_j},v_{\mathrm{jh}}^{{\mathrm{\ψ}}_j}\∈V_J^{{\mathrm{\ψ}}_j};h\∈{\mathrm{\π}}_j^{{\mathrm{\ψ}}_j},{\mathrm{\ψ}}_j\∈\mathrm{\Ψ},j\∈[1,L]\}---\left(13\right)$

Wherein, P ^SThe presentation class scheme, W ^SAnd V ^SRepresent P respectively ^SIn Fuzzy ARTMAP network weight set and corresponding with it mode top weights set, W _j ^SAnd V _j ^SRepresent W respectively ^SAnd V ^SIn belong to the subclass of classification j, w _Jh ^SAnd v _Jh ^SRepresent W respectively _j ^SAnd V _j ^SIn belong to the weight vector of subclass h, π _j ^SBelong to the subclass set of classification j in the presentation class scheme, L is referring to formula (4), Ψ and ψ _jReferring to formula (10), W _j ^{ψ j}And V _j ^{ψ j}Represent example ψ respectively _jIn belong to the Fuzzy ARTMAP network weight subclass of classification j and corresponding with it mode top weights subclass, w _Jh ^{ψ j}And v _Jh ^{ψ j}Represent W respectively _j ^{ψ j}And V _j ^{ψ j}In belong to the weight vector of subclass h.

The present invention adopts two stage example correction strategies, and promptly part is supervised cluster and classification model adjustment

1. part is supervised cluster.Center weight vector V with classification schemes ^S(referring to formula (11)) are as the initial cluster center of a Fuzzy c-means sorter:

$Z=\left\{Z_r\right|Z_r^{\left(0\right)}=v_{\mathrm{jh}}^S;h\∈{\mathrm{\π}}_j^S,j\∈[1,L],r\∈[1,N^S]\}---\left(14\right)$

Wherein, Z represents the cluster centre vector of Fuzzy c-means sorter, z _rR cluster centre of expression Fuzzy c-means sorter, the cluster centre of attention Fuzzy c-means sorter can corresponding independently subclass, N ^SExpression V ^SThe center weights number that comprises, v _Jh ^S, π _j ^SReferring to formula (11), L is referring to formula (4).

This initialization form shown in the formula (14) can guarantee all corresponding definite classification ownership of each cluster centre of Fuzzy c-means sorter, and makes the cluster result of Fuzzy c-means have clear and definite classification information.

Utilize the Fuzzy c-means sorter after the initialization that the sample data collection in the new example is carried out cluster operation.The classification center vector that data characteristics is optimized and revised in the new example of basis can be obtained after the convergence of Fuzzyc-means sorter, the preliminary classification result of sample data in the new example can be obtained simultaneously.

2. classification model adjustment.Cluster centre of setting up according to formula (14) and the corresponding relation between the classification (or subclass), the cluster centre assignment after the Fuzzy c-means convergence is given the center weight vector of classification schemes:

$V^S=\left\{v_{\mathrm{jh}}^S\right|v_{\mathrm{jh}}^S=Z_r^{\left(n\right)};h\∈{\mathrm{\π}}_j^S,j\∈[1,L],r\∈[1,N^S]\}---\left(15\right)$

Wherein, V ^S, v _Jh ^S, π _j ^SReferring to formula (11), L is referring to formula (4), N ^SReferring to formula (14).

Utilize the preliminary classification result of sample data in the new example to train Fuzzy ARTMAP network weight W in the classification schemes again ^S, so far obtained revised classification schemes.With the sorter of the Fuzzy ARTMAP network after training again as unfiled image.Fig. 5 is the classification results figure that utilizes the present invention to obtain.

4. case library upgrades

The relevant information of storing new image with the form of example, with classified sample data and revised classification schemes in the space-time characteristic of new example, the new example as three chief components of storage paradigms:

C ^R＝(Φ ^new，D ^′new，P ^′S) (16)

Wherein, C ^RThe example that the expression case library will be stored, Φ ^NewThe space-time characteristic of representing new example, D ^{' new}The classified sample data of representing new example, P ^{S '}Represent revised classification schemes.With example C ^RStore in the case library reference into as the work of classifying next time.

Utilize SVM method and Cascade sorting technique that unfiled remote sensing image is carried out sort operation respectively, the training sample of two sorters adopts the sample of existing remote sensing image (remote sensing image as shown in table 1).Fig. 3 is the classification results figure that utilizes the SVM method, and Fig. 4 is the figure as a result that utilizes the Cascade sorting technique.Table 3 is that above-mentioned two kinds of sorting techniques and nicety of grading of the present invention compare.Can see that from the result the present invention can effectively avoid the problem of obscuring of atural object classification, improve the nicety of grading of remote sensing image significantly, and can fully excavate and integrate existing remote sensing sample and classificating knowledge.

Table 3 SVM method, Cascade method and nicety of grading of the present invention are relatively

The atural object classification	Nicety of grading
				The SVM method	The Cascade method	The present invention
	Nonirrigated farmland, reed grass glade, water body alkaline land	0.77 0.79 0.72 0.82 0.63 0.8	0.81 0.83 0.7 0.84 0.79 0.82				0.94 0.96 0.87 0.94 0.89 0.88
Overall accuracy				0.76	0.8	0.91

Claims (1)

1. case-based reasoning and the classification of remote-sensing images method that Fuzzy ARTMAP network combines is characterized in that concrete steps are as follows:

(1) make up new example: with the space-time characteristic of unfiled remote sensing image, i.e. coverage and acquisition time are as the space-time characteristic of new example; The sample data set of unfiled remote sensing image is gathered as the sample data of new example.Obtain reflecting the new example of unfiled remote sensing image feature through aforesaid operations.

(2) carry out the space-time characteristic retrieval: the space-time characteristic with new example is a foundation, utilize the space-time similarity function to carry out the space-time characteristic search operaqtion, retrieve in the case library and all satisfying the example set of setting threshold requirement aspect coverage and the acquisition time two with new example from the coverage coincidence degree of remote sensing image, degree of approximation two aspects of remote sensing image acquisition time.

(3) it is selected to carry out content characteristic: utilize the classification model in the example that preliminary search goes out that new example sample data is carried out sort operation.The corresponding classification results that obtains with each example is a foundation, and the center weight vector of utilization calculates the retrieval energy function of each classification of each example correspondence.Each classification is asked for the example index value with minimum retrieval energy function, set up paradigm index set after selected with this.

(4) be fundamental construction preliminary classification scheme with the paradigm index set after selected: the classification model of respective classes in the example of index value correspondence is extracted, be reassembled as the preliminary classification scheme.

(5) be foundation with the data characteristics in the new example, revise the preliminary classification scheme:

1. with of all categories in the classification model or subclass the center weights as the preliminary examination cluster centre of Fuzzy c-means sorter, can guarantee that like this each cluster centre can both corresponding definite classification logotype.

2. according to Fuzzy c-means iterative algorithm the data in the new example are carried out cluster operation, obtain cluster centre after the algorithm convergence, also obtain the preliminary classification result of new example data simultaneously according to the of all categories or subclass of new example data feature adjustment.

3. the cluster centre of the Fuzzy c-means after will restraining is as the mode top weights of classification schemes; The network weight that utilizes the preliminary classification result of new example data to train Fuzzy ARTMAP in the classification schemes again.So far obtain revised classification of remote-sensing images scheme.

(6) utilize the Fuzzy ARTMAP network weight in the classification of remote-sensing images scheme that remote sensing image is carried out sort operation, sorting algorithm is according to the sorting algorithm of Fuzzy ARTMAP network.Obtain the classification of remote-sensing images result.

(7) with the space-time characteristic of new example, revised classification of remote-sensing images scheme and sorted remote sensing image data chief component as the example that will store, this example is deposited in the case library in order to sort operation reference next time, finish the renewal operation of case library.

Images