Distinguishing Weather Phenomena From Bird Migration Patterns in Radar Imagery

Aruni RoyChowdhury, Daniel Sheldon, Subhransu Maji, Erik Learned-Miller; Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2016, pp. 10-17

Data archived by the United States radar network for weather surveillance is useful in studying ecological phenomena such as the migration patterns of birds. However, all such methods require a manual screening stage from domain experts to eliminate radar signatures of weather phenomena, since the radar beam picks up both biological and non-biological targets. Automating this screening step would be of significant help to the large scale study of ecological phenomenon from radar data. We apply several techniques to this novel task, comparing the performance of Convolutional Neural Networks (CNNs) models against a baseline of the Fisher Vector model on SIFT descriptors. We compare the performance of deeper and shallower network architectures, deep texture models versus the regular CNN model and the effect of fine-tuning ImageNet pre-trained networks on radar imagery. Fine-tuning the networks on the radar imagery provides a significant boost, and we achieve a final accuracy of 94.4%.