Exploration of Resnet Variants in High Spatial Resolution Domain Adaptation

From air-to-space imagery

Abstract

Land cover is nowadays mapped mostly from airborne and space-borne data. Because of the difference in sensors, large spectral differences and inconsistent spatial resolution may arise between these two data sources. Consequently, the same object may exhibit completely different features. In this case, models trained from annotated airborne and ineffective when applied to space-borne data. Cross-Sensor Land-COVER (LoveCS) shows good results in overcoming this problem. LoveCS leverages small-scale aerial image annotations to promote land cover mapping on large-scale spacecraft. LoveCS uses ResNet50 as its encoder. In recent years, many studies have tried to develop other variants of ResNet, such as ResNeXt, ResNeSt, Res2Net, and Res2NeXt. These variants turned out to give better results in a variety of tasks compared to ResNet. Therefore, in this study we modified the LoveCS encoder by replacing ResNet50 with ResNet variants such as ResNeXt, ResNeSt, Res2Net, and Res2NeXt in an effort to improve LoveCS accuracy. Our evaluation shows that Res2Net50 as an encoder improves the performance of LoveCS. The average F1 increases by 1.38%, OA by 1.96%, and Kappa by 2.75% from the baseline method.