Three-Dimensional Convolutional Neural Network Algorithm For Forest Tree Species Classification Using Airborne Hyperspectral Spatial-Spectral Joint Features

To explore the potential of deep convolutional neural networks(CNN)in airborne hyperspectral data classification,so as to improve the classification accuracy of forest tree species,the aeronautical hyperspectral data of Nanning Gaofeng Forest Farm in Guangxi province obtained by the LiCHy system of Chinese Academy of Forestry was used.The CNN model proposed in this paper aims to deal with hyperspectral image analysis in an end-to-end manner.It can use raw data as input,without dimension reduction or feature screening,thus eliminating the need for traditional classification methods to manually feature selection in different degrees.The 3D convolutional layers in the network can extract spectral and spatial features simultaneously,learn the local signal changes in the spatial and spectral dimensions of the feature cube,and classify them with important recognition features to improve the discriminating ability of hyperspectral images.For the issues that the airborne hyperspectral data dimension is high and the training samples are relatively small,the CNN model is optimized to make full use of the spectral and spatial information of the original data,while avoiding over-fitting caused by small sample training.Compared with the traditional feature selection and object-oriented segmentation methods,the proposed three dimentional CNN can obtain higher classification accuracy with the overall accuracy of 98.38%,Kappa coefficient of 0.98.Compared with support vector machine(SVM)combined with random forest(RF)feature selection classification,the overall accuracy is improved by 8.82%,and the Kappa coefficient is increased by 0.11.In the case of small sample training(75%reduction in training samples),the overall accuracy can still reach 95.89%,and the Kappa coefficient is 0.94.The three-dimensional convolutional neural network can fully utilize the rich information in the feature extraction and classification of airborne hyperspectral imagery,which can distinguish the southern forest species very effectively;in addition,the reasonable network structure and training strategy(Joining the Dropout layers)can strongly accelerated the network training speed and still get good results in small sample training.By this method,efficient and accurate forest species classification can be achieved.