On Cloud N
Computer Vision
Segmentation
UNet
fastai
Competition
Segment Satelite data into Cloud and No Cloud.
This Challenge was hosted on DrivenData and the goal is to segment clouds in Sentinel-2 Satelite data. Besides visual image data, the Sentinel-2 dataset also offers different non-visual bands, in particular infrared. This solution uses geometric augmentations on all four channels and “visual” augmentations on the rgb-channels only. The model is a simple U-Net that is pretrained on 440x440 random crops and fine tuned on the whole image in the original chip resolution (512x512).