2020/09/30 by Maxim Neumann, André Susano Pinto, Neumann, Maxim +5 · 2 citations
Computer Science · Engineering · #Advanced Image and Video Retrieval Techniques #Computer Vision and Pattern Recognition (cs.CV) #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences #Machine Learning (cs.LG) #Remote-Sensing Image Classification
paper · pdf · doi:10.48550/arxiv.2010.00332
openalex publication_date 2020/09/30 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
Automatically finding good and general remote sensing representations allows\nto perform transfer learning on a wide range of applications - improving the\naccuracy and reducing the required number of training samples. This paper\ninvestigates development of generic remote sensing representations, and\nexplores which characteristics are important for a dataset to be a good source\nfor representation learning. For this analysis, five diverse remote sensing\ndatasets are selected and used for both, disjoint upstream representation\nlearning and downstream model training and evaluation. A common evaluation\nprotocol is used to establish baselines for these datasets that achieve\nstate-of-the-art performance. As the results indicate, especially with a low\nnumber of available training samples a significant performance enhancement can\nbe observed when including additionally in-domain data in comparison to\ntraining models from scratch or fine-tuning only on ImageNet (up to 11% and\n40%, respectively, at 100 training samples). All datasets and pretrained\nrepresentation models are published online.\n