A deep learning approach for safeguards-relevant change detection using Sentinel-2 imagery

Lisa Beumer - Forschungszentrum Julich GmbH, Institute of Energy and Climate Research, Nuclear Waste Management (IEK-6), Germany
Irmgard Niemeyer - Forschungszentrum Julich GmbH

Satellite imagery represents a key source of information for the implementation and verification of nuclear non-proliferation, arms control and disarmament treaties. In particular, change detection has become an important application of remote sensing and also has potential for nuclear verification. In recent decades, deep learning algorithms have achieved several stunning achievements. Most notably, convolutional neural networks (CNNs) have demonstrated their image understanding capabilities. Even though huge amounts of remote sensing data exist, most of the data is unlabeled and thus inaccessible to supervised learning algorithms. Transfer learning approaches can reduce the data requirements of deep learning algorithms because the already pre-trained machine learning model can be built with comparatively little training data. Having freely available large-scale training datasets is key to these advances. ImageNet, for example, has more than 14 million tagged images and serves as a benchmark for large-scale object recognition and image classification systems [1]. However, while most of the available pre-trained models are based on ImageNet, their application to specific remote sensing applications, such as nuclear verification, is not guaranteed. Moreover, this dataset is limited to the RGB color space. In contrast, remote sensing has long taken advantage of multispectral images. In this work, we propose a method to use unlabeled data for pre-training remote sensing representations and deploy a CNN that can effectively extract contextual information at multiple levels by using all bands in multispectral images, which are in turn used for change detection. Three state-of-the-art CNNs, respectively Residual Neural Network (ResNet) [2], AlexNet [3], and Dense Convolutional Network (DenseNet) [4], are evaluated towards fine-tuning and deep CNN feature extraction. We perform experiments with the EuroSAT dataset, which is based on Sentinel-2 satellite imagery covering 13 spectral bands and consisting of 10 classes with 27,000 labeled and georeferenced patterns [5], and the BigEarthNet dataset, a large-scale Sentinel-2 benchmark archive consisting of 590,326 Sentinel-2 image patches, as training datasets [6]. As test data, we focus on Sentinel-2 images provided by the Copernicus Open Access Hub.