Transfer Learning with Synthetic Satellite Imagery
Victor Manuel Vergara, Jeremy J. Wojcik, Francisco O. Viramontes, Mari A. Aoki, Evan T. Kain, Tyler M. Lovelly · 2024
The field of satellite imagery suffers from scarce availability of open datasets that can be used to develop novel algorithms. One of the most recent open data sets, the RarePlanes dataset, provides real satellite images with an excellent resolution and hand-made annotations. The RarePlanes dataset provides satellite images of real aircraft parked along runways. In the context of training deep convolutional neural networks (CNNs), the RarePlanes dataset has a class imbalance where some aircraft classes are sufficiently represented, and others suffer from a short supply of annotated instances. With this pitfall in mind, the RarePlanes dataset included synthetic data that can be used to compensate for problems raised during CNN training. This report assesses the use of synthetic satellite imagery to improve CNN training of real satellite images using the transfer learning (TL) technique. TL with synthetic satellite imagery is compared against TL with Common Objects in COntext (COCO) dataset and against the case of no- TL with randomly initialized weights. Results indicate that TL with synthetic satellite imagery provides better results when applied to real satellite imagery supporting the use of synthetic data for real data in CNN applications.