Pixel-wise Classification of Hyperspectral Images with 1D Convolutional SVM Networks

Nowadays, remote sensing image analysis is needed in various important tasks such as city planning, land-use classification, agriculture monitoring, military surveillance, and many other applications. In this context, hyperspectral images can play a useful role, but require specific handling. This paper presents a convolutional neural network based on one-dimensional support vector machine (SVM) convolution operations (1D-CSVM) for the analysis of hyperspectral images. SVM-based CNN (CSVM) was introduced first for the classification of high spatial resolution RGB images. It relies on linear SVMs to create filter banks in the convolution layers. In this work, the network is modified to cope with one-dimensional hyperspectral signatures and perform pixel-based classification. It thus analyzes each pixel spectrum independently from the pixel spatial neighborhood. Experiments were carried out on four benchmark hyperspectral datasets, Salinas-A, Kennedy Space Center (KSC), Indian Pines (IP) and Pavia University (Pavia-U). Compared to state-of-the-art models, the proposed network produces promising results for all tested datasets, with an accuracy up to 99.76%.