Learning Robust Deep Features for Efficient Classification of UAV Imagery

摘要

This paper presents a deep learning approach for the classification of Unmanned Aerial Vehicle (UAV) images acquired by different sensors and different locations of the earth surface. In a first step, the labeled and unlabeled UAV images under analysis are fed to a pretrained convolutional neural network (CNN) for generating initial deep feature representations. Then in second step, we learn robust domain-invariant features using an additional network composed of two fully connected layers. This network aims to tackle the data-shift problem by reducing the discrepancy between the labeled and unlabeled data distributions. For such purpose, the first layer projects the labeled data to the space of the unlabeled data, while the second layer maintains the discrimination ability between the different land-cover classes. Experimental results obtained on two datasets acquired over the cities of Trento and Toronto with spatial resolutions of 2 cm and 15 cm, respectively, are reported and discussed.