Pixel-Based Classification of Hyperspectral Images Using Convolutional Neural Networks

The recent progress in geographical information systems, remote sensing (RS) and data analytics enables us to acquire and process large amount of Earth observation data. Convolutional neural networks (CNN) are being used frequently in classification of multi-dimensional images with high accuracy. In this paper, we test CNNs for the classification of hyperspectral RS data. Our proposed CNN is a multi-layered neural network architecture, which is tailored to classify objects based on pixel-wise spatial information using spectral bands of hyperspectral imagery (HSI). We use benchmark satellite imagery in four different HSI datasets for classification using the proposed architecture. Our results are compared with support vector machine (SVM) and extreme learning machine (ELM) algorithms, which are frequently used techniques of machine learning in RS data classification. Moreover, we also provide a comparison with the state-of-the-art CNN approaches, which have been used for HSI classification. Our results show improvements of up to 6% on average over SVM and ELM while up to 4% improvement is observed in comparison with two recently proposed CNN architectures for HSI classification accuracy. On the other hand, the processing time of our proposed CNN is also significantly lower.