Multi-Instance Neural Network Architecture for Scene Classification in Remote Sensing

Scene classification is important problem in remote sensing since it is prerequisite to other more intelligent analysis operations. Often times, for a given scene only one part of it indicates which class it belongs to, whereas the other parts are either irrelevant or they actually tend to another class. To address this problem, we propose to divide the RS scene into multiple sub-images, consider each one as an instance of the original scene. In that case, we can view the scene as a bag of instances having the same label and we proposed a deep multi-instance learning (MIL) architecture for the classification problem. Our proposed deep MIL architecture extracts CNN features from each instance, then learns how to fuse them into one feature using weighted average layer in the network. The weights used in this layer are automatically learned by the network for each scene. We test the proposed multitask network on three popular scene datasets, namely UC Merced, KSA, and AID datasets. Preliminary results show the promising capabilities of this solution at improving the classification accuracy by giving more weight to the feature extracted from the more informative scene instance.