Deep Feature Encode Based Hyper Spectral Classification

With the application of hyperspectral image data more widely, and hyperspectral image data fusion of spectral information and geographical information surface material, the classification has been the focus and difficulty of hyperspectral image research.Firstly, this paper describes the characteristics and current classification methods in hyperspectral image, including several types of common classification. Summarizes some of the current methods of hyperspectral image preprocessing classification dimensionality reduction, and gives the definition of assessment and classification algorithm performance indicators Kappa coefficient and classification accuracy.Secondly, the introduction of the neural network classifier a theoretical foundation and model-based neural network stack to build self-encoding algorithm is given. Basing on the fusion of hyperspectral image data and Stack AutoEncode algorithm(SAE), 2 hyperspectral classification algorithm Deep Future Encode Classification(DFE) and Deep Future Encode Classification based Support Vector Machine(DFESVM) are built in this paper.And numerical verification is established in the database KSC and Pavia. DFE and DFESVM Classification accuracy(96.1%, 96.5%, KSC; 93%, 92.3%, Pavia) and kappa coefficient(95.6%, 96.1%, KSC; 93%, 91.8%, Pavia), higher than the SVM and SAE 3%.After some analysis of numerical experiment, we get that Classification accuracy and kappa coefficient as the hidden layer unit number of different changes, when it gets 60, classification is the best. With the increase of network training parameters epochs, the classification accuracy and kappa coefficient is higher. When the sample proportion of training is small, and its impact on the classification accuracy and kappa coefficient is large, the greater the ratio, the higher classification accuracy and kappa coefficient will be got, which can reach 97.3%(Pavia).Finally, this thesis summarizes the work carried out.