Validation Of Planting Structure In The Middle Reaches Of Heihe River

With advantages of narrow band, abundant information, Hyperspectral image has awide applications in fields of geology, agriculture, environment,military,hydrological,atmosphere and so on. Hyerspectral images of CASI and SASI were finely classified inorder to explore application potentialities in the plant structure classification, to testreaction ability of the ground truth of CASI qualitative and to validate ASTER data.Zhangye City has been one of the ten largest commodity agriculture area in China, withseed corn as an important economic crop, so the research done in this paper has greatsignificance in the follow-up agriculture and crop estimation and detection.Maximum likelihood,Artificial Neural Network and Support vector machineclassifiers were adopted as classification methods in this paper. The CASI and SASIimages were strictly pre-processed by using the field spectral measurement data(ASD)and CE-318data.(1) There are great maximum likelihood classification differences between CASI andCASI+SASI, Overall accuracy and kappa coefficients were93.8642%,56.9114%and0.9365,0.5043respectively. Under limited samples, maximum likelihoodclassifier is less efficient to classify CASI+SASI data, but it was suit for CASI.(2) When Training Rate, Training Momentum, Training RMS Exit Criteria and Numberof Training Iterations were not changed, the NN were used to classify CASI andCASI+SASI just by adjusting Number of Hidden Layers, Training ThresholdContribution. Results indicated that convergence speed became slow and weights ofTraining Threshold Contribution decreased from0.99to0.93with the increase ofNumber of Hidden Layers. No rules could be found in classification accuracy withthe increase of Training Threshold Contribution. The best classification accuracy ofCASI was92.9679%, and CASI+SASI was90.2447%.(3) Kernel functions and penalty factors were two main factors to affect the effects ofsupport vector machine (SVM). Overall accuracy was rising with the increasing ofpenalty factors, but limited. There were some differences in results with same penaltyfactors but different kernel functions. On the contrast, the result was less differences.SVM was barely influenced by hughes. (4) Experience combination of classification results was like the average singleclassification, was also seen as the optimization of single classification results, whichbased on the premise that group judgment reliability is higher than individualjudgement s, and the combination of classification results also depended on thetraining recognition rate of sample. The combination of classification resultstheoretically could improve precision generally, but not necessarily better than thebest classification results of single classifiers.(5) PCA and MNF transformation reflected different characteristics of information bycompressing data and extracting information from different sides of the data. Themaximum likelihood classification accuracy after MNF transformation was96.52%,better than maximum likelihood, neural network, support vector machine (SVM)classification and combination of three classifiers.According to classification results, ASTER classification, classification of averagepixel up-scaling data of CASI and classification of interpolation up-scaling of CASI werevalidated by maximum likelihood classification after the MNF transformation of CASI.Results indicated that12,6,3,3categories were got by CASI data, ASTER data, averagepixel scale CASI and interpolation up-scaling of CASI data respectively. Up-scaling dataof CASI could not be enough for ground truth data. Experience combination of classeswas the best way to covert scale. Compared with CASI classification, ASTERclassification was less detail. Correct classification of ASTER was56.394%withexperience combination classes of CASI classification as ground truth date.