Document Type : Original Article
School of Surveying and Geospatial Engineering, College of Engineering, University of Tehran
In recent years, deep learning methods based on the convolutional neural network (CNNs) have demonstrated good performance for hyperspectral image classification (HSI). Although, in order to obtain good results, we need a large number of training data in the CNNs to avoid the overfitting problem. This paper aims to establish a segmentation-based method to extend the training data for deep learning-based hyperspectral image classification. First, two unsupervised segmentation methods (K-Means and Multi-resolution) are used for the segmentation of the hyperspectral images. Second, we obtained pseudo-training data which depends on the overlay between segmented hyperspectral images and original training data sets. So we extend the number of training samples for CNN to avoid the overfitting problem and achieve good results. Finally, a Hybrid-CNN model that is a combination of 2D-Convolution and 3D-Convolution is applied to classify hyperspectral datasets with the training samples consisting of the original and pseudo training sets. The proposed method was tested on two Kennedy Space Center (KSC) and Botswana hyperspectral images and the results are compared with the two methods. The overall accuracy with the proposed method retrieves 100% and 96.11% for KSC and Botswana datasets, respectively. Also, we tested the proposed Hybrid-CNN network with Pavia University data, and the classification results show that the proposed Hybrid-CNN has good performance in the face of complex data. The overall accuracy retrieves 99.66% for the Pavia dataset. Keywords: Hyperspectral Image Classification (HSI), Convolutional Neural Network (CNN), Multiresolution segmentation, K-Means Clustering.