Analysis of learned features for remote sensing image classification

Convolutional neural networks (convnets) have shown excellent results in various image classification tasks. Part of the success can be attributed to good image representations that are extracted using convolutional layers of the network. In this paper we consider convnets from the perspective of feature extraction for remote sensing image classification. We analyze the impact of convolutional feature extraction as well as the role of feature learning on the ability of features to discriminate between land cover classes. The quantitative analysis is based on measuring both classification accuracy and discriminative ability of features. For the latter we use Fisher discriminant analysis and show that features extracted using convolutional layers with random weights have significant discriminative ability and result in a reasonable baseline for remote sensing image classification, which suggests that convolutional feature extraction itself is an important ingredient of feature extraction in convnets. Using learned convnets for feature extraction further improves discriminative ability of features.