Automatic discriminative feature extraction using Convolutional Neural Network for remote sensing image classification

Abstract

Supervised approaches require selection of training samples from all classes and then a set of highly discriminative feature descriptors are extracted to represent each class. Traditional classification methods employee handcrafted features. Although, such features are effective but usually require prior knowledge and involve lot of laborious work. In addition, the availability of less training samples for multi/hyperspectral data makes the problem even more challenging. An alternative solution could be to employee deep learning-based approach for automatic extraction of highly stable feature patterns from input data. This paper proposes a new method using a deep learning based on Convolutional Neural Network (CNN) for automatic extraction of spectral-spatial features from high-resolution multi-spectral images. The learning framework consisted of a series of convolutions and pooling layers. We evaluated the effectiveness of the proposed method for the problem of land cover classification. The dataset consisted of 10 high-resolution multi-spectral images obtained from Rize district of Turkey. The classification was then performed by applying the random forest classifier. The results indicated that the proposed method was effective and easier to implement and learn. © 2020 40th Asian Conference on Remote Sensing, ACRS 2019: "Progress of Remote Sensing Technology for Smart Future". All rights reserved.