Aerial-CAM: Salient Structures and Textures in Network Class Activation Maps of Aerial Imagery

摘要

This paper aims at visualizing how deep networks interpret aerial scenes by examining their internal representations. We utilize Class Activation Mapping (CAM) techniques to obtain a view of a deep network's perception of aerial images and identify salient local regions. We apply our methods on two remote sensing datasets, the AID dataset and the UC Merced Land use dataset, and we show that local structures and textures emerge in the most active regions of aerial images. We then analyze these interpretations when the network is trained on one dataset and tested on another to demonstrate the robustness of feature learning across aerial datasets. We finally visualize these interpretations when transfer learning is performed from an aerial dataset (AID) to a generic object dataset (MS-COCO) to illustrate how transfer learning benefits the network's internal representations.