ECoNets: Rotation Equivariant Contrail Detection Neural Networks in Satellite Imagery

Edgar Loza, Davide Di Giusto, Vincent R. Meijer, Teodora Petrisor

Keywords: Contrails, Image Segmentation, $C_N$- and $D_N$-Equivariance

Abstract: We present ECoNets, equivariant U-Net models applied to contrails segmentation in satellite images. In the context of a highly class-imbalanced problem with scarce annotated data, equivariant models benefit from higher segmentation scores and faster convergence, while requiring fewer trainable parameters, in all settings and in particular in a reduced training dataset size regime. We benchmark ECoNets on the OpenContrails satellite imagery dataset as well as on a smaller in-house labelled dataset of Meteosat Second Generation (MSG) geostationary satellite images in order to assess fine-tuning equivariant models for contrail detection over Europe.

Submission Number: 63