Introduction
This is my graduate thesis, whose topic is [Hyperspectral Image Classification Using Deep Neural Network]. I would like to express my sincere gratitude to my supervisor, Professor PHILIP HENG WAI LEONG, for his support and encouragement of my work. Without his help and expert advice, it is impossible to complete this thesis.
Extract
Hyperspectral image is a commonly used technique for earth surface survey in ecology, mining, and hydrological application. Compared with other survey methods, it has the advantage of lower cost and faster data collection. Briefly, it is an image form that contains hundreds of narrow spectral channels in every single pixel, which is a measured value of the corresponding wavelength. By processing the spectral and spatial informa- tion in the hyperspectral image, each pixel and the surface objects they represent will be identified and classified. At present, the most commonly used method for processing hyperspectral images is based on deep learning. However, due to some reasons, such as the defects of neural network design and fewer available training samples, the per- formance of classification needs to be improved. My main contribution in this thesis is that: 1)A preprocessing of the presentation extraction on hyperspectral image data set is implemented, which is utilized to e ciently extract the spatial and spectral in- formation of hyperspectral images. Propose an improved extraction technique on edge pixels, which includes more accurate spatial information. 2) Propose a deep convolu- tional neural network model for hyperspectral image classification. Many techniques, such as dropout and regularization, are applied to optimize the neural network perfor- mance, solving problems like overfitting, and the small mount of training samples. 3) Experiments are conducted on actual hyperspectral image data set, and compared with other peoples previous work. A high classification accuracy is obtained as the result.
View in PDF
For reference only, plagiarism is forbidden
See Thesis in pdf
