Land Cover Classification from Satellite Imagery with U-Net and Lovász-Softmax Loss

The land cover classification task of the DeepGlobe Challenge presents significant obstacles even to state of the art segmentation models due to a small amount of data, incomplete and sometimes incorrect labeling, and highly imbalanced classes. In this work, we show an approach based on the U-Net architecture with the Lov́asz-Softmax loss that successfully alleviates these problems; we compare several different convolutional architectures for U-Net encoders.