Automated Machine Learning for Satellite Data: Integrating Remote Sensing Pre-trained Models into AutoML Systems

Abstract

Current AutoML systems have been benchmarked with traditional natural image datasets. Differences between satellite images and natural images (e.g., bit-wise resolution, the number, and type of spectral bands) and lack of labeled satellite images for training models, pose open questions about the applicability of current AutoML systems on satellite data. In this paper, we demonstrate how AutoML can be leveraged for classification tasks on satellite data. Specifically, we deploy the Auto-Keras system for image classification tasks and create two new variants, IMG-AK and RS-AK, for satellite image classification that respectively incorporate transfer learning using models pre-trained with (i) natural images (using ImageNet) and (ii) remote sensing datasets. For evaluation, we compared the performance of these variants against manually designed architectures on a benchmark set of 7 satellite datasets. Our results show that in 71% of the cases the AutoML systems outperformed the best previously proposed model, highlighting the usefulness of a customized satellite data search space in AutoML systems. Our RS-AK variant performed better than IMG-AK for small datasets with a limited amount of training data. Furthermore, it found the best automated model for the datasets composed of near-infrared, green, and red bands.