Improved Convolutional Neural Networks for Hyperspectral Image Classification

Classification is the process of setting class labels to pixels based on some obtained properties. Hyperspectral images (HSI) have very high dimensionality, which results in higher cost and complexity for analyzing and classifying them as superfast processors and large storage devices are required. Moreover, due to limited training samples and labeled data, classification remains an arduous task. Many methods have been presented till now for classification of HSI based on traditional methods that use handcrafted features beforehand, principal component analysis and its variations, decision trees, random forests, SVM-based methods, and neural networks, but most of these consider only the spectral information for classification resulting in low classification accuracy. Nowadays, increasing spatial resolution of HSI demands obtaining spatial data for further improving classification performance. We, therefore, present a classification method which obtains spectral as well as spatial features using convolutional neural network (CNN) model and then a logistic regression (LR) classifier that uses the activation function softmax for predicting classification results. Our proposed method is compared with considered techniques and tests on HSIs, namely, Indian pines and Pavia University, which have shown better performance regarding parameters such as overall accuracy (OA), average accuracy (AA), and kappa coefficient (K).