Research On Lightweight Hyperspectral Image Classification

The spectral imager simultaneously images the target area with dozens to hundreds of continuous and subdivided spectral bands to form a hyperspectral remote sensing image.In addition to the spatial information such as the structure,position relation and shape of the ground object,the spectral information which can represent the physical material of the object is also obtained.The advantages of "atlas integration" make hyperspectral remote sensing images play an important role in environmental monitoring,urban planning,smart agriculture and other fields.The fine classification of ground objects using hyperspectral images is the core content of the application of hyperspectral remote sensing technology.Some artificial intelligence methods,represented by deep learning,relying on highperformance computing equipment are often used in the classification of hyperspectral remote sensing images,because their classification accuracy is significantly better than traditional learning methods.However,the classification method based on deep learning still has some problems.Deep learning methods fit the training data well by building a deep network and increasing the parameters of the model.However,as the scale of the model increases,the amount of computation in the training process also increases greatly,which not only reduces the model efficiency,but also requires the use of high-performance computing equipment to train the network.Therefore,it does not suitable for some scenarios that that need to consider production costs,such as small and medium-sized farms,which have insufficient equipment resources in the process of developing smart farms.In addition,data labeling of hyperspectral images is an expensive and time-consuming task,while supervised deep learning methods often require a large amount of labeled data to train models.To solve the above problems,this thesis proposes two hyperspectral image classification methods based on convolutional neural network.The main work of this thesis is as follows:（1）A lightweight hyperspectral remote sensing image classification method based on convolutional neural network is proposed,which aims to ensure high classification accuracy while reducing model training costs,so as to help small and medium-sized farms avoid replacing expensive high-performance equipment and reduce operating costs.Data preprocessing methods such as principal component analysis and data augmentation are used to reduce dimensionality and sample expansion of hyperspectral remote sensing images.Spectral-spatial joint features are introduced to improve the classification accuracy.The structure of network is optimized to accelerate the training process.Finally,by conducting experiments on three small and medium-scale benchmark datasets,and comparing with some traditional classification methods and deep learning methods,the results show that this method can guarantee a good classification result and reduce the training cost.（2）A lightweight hyperspectral remote sensing image classification method based on the combination of 3D convolutional neural network and 3D convolutional autoencoder is proposed.By using 3D convolution,the medium parameters of network model are effectively reduced and the joint features of local space spectrum of hyperspectral images are extracted.At the same time,3D convolutional autoencoder is used to pre-train the 3D convolutional neural network with unlabeled samples,and a small number of labeled samples are used to fine-tune the classification model.Experiments show that the proposed method can obtain better classification accuracy while reducing the usage of label sample.