Deep Learning - a New Approach for Multi-Label Scene Classification in Planetscope and Sentinel-2 Imagery

Motivated by the increasing availability of high-resolution satellite imagery, we developed deep learning models able to efficiently and accurately classify the atmospheric conditions and dominant classes of land cover/land use in commercial PlanetScope imagery acquired over the Amazon rainforest. In specific, we trained deep convolutional neural network (CNN) to perform multi-label scene classification of high-resolution (<10 m) satellite imagery. We also discuss the challenges and opportunities in training deep CNN models for multi-label scene classification. Finally, we investigate the transferability of our PlanetScope-trained models to freely available Sentinel-2 imagery acquired over the wet tropics of Australia. Our best performing model achieved an $F$β of 0.91, which was only 2% short of the top performing model in the Understanding the Amazon from Space Kaggle competition [1]. We also find that our models are suitable for classifying similar resolution satellite imagery, such as Sentinel-2.