Detection of construction objects using time series of Sentinel-2 data and neural networks

Abstract:

Land cover and land use information is significant for various geospatial applications, such as urban expansion. It is also a foundation to characterize and analyze the constant changes on the surface of the Earth and associated socio-ecological interactions. The increased number and quality of space-borne very high spatial resolution (VHR) sensors (at 1-10 m), including micro- and nano-satellites, provide new opportunities to exploit multiple VHR data sources for generating new remote-sensing products for the study of land cover and land use change. With higher spatial, spectral and temporal resolution data available, for example, from Sentinel-2 (10 m) or Planet/Dove constellations (3 m), it becomes possible for land monitoring at finer spatial resolution.

This study focuses on the detection of anthropogenic activities, in particular constructions in the built environment (urban expansion). We will address the question that what accuracy can be achieved for detection and persistent monitoring of anthropogenic progressive activities (such as constructions) by applying machine learning techniques to Sentinel-2 imagery. Two benchmark datasets are used: Onera Satellite Change Detection Dataset, which is based on Sentinel-2 data and binary labels (change/no-change), and Washington DC dataset of land cover changes developed by us. We improve the Onera benchmark dataset with further co-registration and atmospheric correction using the well-established LaSRC (Land Surface Reflectance Code) algorithm to match the data used in the Washington DC dataset. We also plan to expand the Onera dataset to include more labels that will incorporate not only the fact of change but also the type of changes. We use Onera dataset to train the deep learning model (U-Net) and transfer the trained network to the Washington DC. In addition to the U-Net, we also explore traditional pixel-based and object-based image analysis approaches. We will show the derived change maps from different methods and evaluation metrices containing the precision, recall and F1 rate from the point of view of the change class.