Convolutional Neural Network for Hyperspectral Image Classification

摘要

Hyperspectral images (HSI) have been used in many domains such as agriculture, healthcare, food security, etc. With the recent development in hyperspectral imaging technology, hyperspectral images'' spectral and spatial resolutions have improved considerably and provided opportunities to characterize objects on the Earth''s surface with high precision.　　With the progress of deep learning techniques in computer vision, convolutional neural networks (CNNs) have been widely used for hyperspectral image classification. However, most of these methods are inefficient in real-world application due to the high dimensionality of the hyperspectral data, lack of large-labeled datasets, and spatial information integration. In this thesis, several convolutional neural networks have been explored and optimized for hyperspectral image classification, including crop classification.　　Our first approach focuses on combining a 3D-2D convolutional neural network (CNN) in an end-to-end manner to mitigate the long-training times of hyperspectral images. For this purpose, a fast-learning block based on 3D depth-wise convolution layer, a dimension reduction block followed by a 2D convolutional neural network was introduced to extract spectral-spatial features. The experiments on four datasets showed that the proposed approach increases the classification performance and can significantly reduce the network parameters compared to some state-of-art methods. To our knowledge, this approach was the first attempt in the literature to use an end-to-end 3D-2D convolutional neural network for HSI classification. It was also the first-time using 3D depthwise separable convolutional neural network in HSI classification.　　Our second approach proposes an extended morphological profile cube (EMPC) and a new densenet to tackle the limited training samples issue of hyperspectral image classification. In addition, the 3D densely connected network uses spectral-spatial dense connectivity blocks for early features fusion and the dimension reduction blocks for reducing feature maps dimensions and the number of network parameters. Furthermore, the Minimum Noise Fraction transform was applied to reduce the high spectral dimension of HSI. Extensive experiments proved the robustness of the proposed approach using limited training samples over some state-of-art methods. To our knowledge, this was theamp;nbsp;first attempt in the literature to classify HSI using features extracted with EMPC and 3D CNN.　　Our last approach proposes a new 3D CNN with covariance pooling for spectral-spatial UAV-borne HSI crop classification. Specifically, an efficient learning module, a context guided block and covariance pooling are proposed to integrate the spectral-spatial context information efficiently and effectively. Extensive experiments on Honghu and LongKou datasets showed that the proposed approach has at least 1.5 × fewer parameters and can achieve 99.57%and 99.93%of overall accuracy over both datasets, which outperforms some state-of-art CNN-based Unmanned Aerial Vehicle (UAV)-borne HSI crop classification methods.　　The results of these methods proved the feasibility of improving the CNNs based methods with less parameters for HSI classification, which can be used as an efficient pipeline for HSI classification in real world-application.

目录