An ensemble learning approach for the classification of remote sensing scenes based on covariance pooling of CNN features

This paper aims at presenting a novel ensemble learning approach based on the concept of covariance pooling of CNN features issued from a pretrained model. Starting from a supervised classification algorithm, named multilayer stacked covariance pooling (MSCP), which exploits simultaneously second order statistics and deep learning features, we propose an alternative strategy which employs an ensemble learning approach among the stacked convolutional feature maps. The aggregation of multiple learning algorithm decisions, produced by different stacked subsets, permits to obtain a better predictive classification performance. An application for the classification of large scale remote sensing images is next proposed. The experimental results, conducted on two challenging datasets, namely UC Merced and AID datasets, improve the classification accuracy while maintaining a low computation time. This confirms, besides the interest of exploiting second order statistics, the benefit of adopting an ensemble learning approach.