Leveraging Bioclimatic Context for Supervised and Self-Supervised Land Cover Classification

Modern neural networks achieve state-of-the-art results on land cover classification from satellite imagery, as is the case for almost all vision tasks. One of the main challenges in this context is dealing with geographic variability in both image and label distributions. To tackle this problem, we study the effectiveness of incorporating bioclimatic information into neural network training and prediction. Such auxiliary data can easily be extracted from freely available rasters at satellite images’ georeferenced locations. We compare two methods of incorporation, learned embeddings and conditional batch normalization, to a bioclimate-agnostic baseline ResNet18. In our experiments on the EuroSAT and BigEarthNet datasets, we find that especially the use of conditional batch normalization improves the network’s overall accuracy, generalizability, as well as training efficiency, in both a supervised and a self-supervised learning setup. Code and data are publicly available at https://t.ly/NDQFF.