Neural Network-Based Urban Change Monitoring with Deep-Temporal Multispectral and SAR Remote Sensing Data

• Main Authors: Georg Zitzlsberger, Michal Podhorányi, Václav Svatoň, Milan Lazecký, Jan Martinovič
• Format: Article
• Language: English
• Published: MDPI AG 2021-07-01
• Series: Remote Sensing
• Subjects: urban change detection; continuous urban monitoring; neural network; SAR; optical multispectral; deep-temporal
• Online Access: https://www.mdpi.com/2072-4292/13/15/3000

Remote-sensing-driven urban change detection has been studied in many ways for decades for a wide field of applications, such as understanding socio-economic impacts, identifying new settlements, or analyzing trends of urban sprawl. Such kinds of analyses are usually carried out manually by selecting high-quality samples that binds them to small-scale scenarios, either temporarily limited or with low spatial or temporal resolution. We propose a fully automated method that uses a large amount of available remote sensing observations for a selected period without the need to manually select samples. This enables continuous urban monitoring in a fully automated process. Furthermore, we combine multispectral optical and synthetic aperture radar (SAR) data from two eras as two mission pairs with synthetic labeling to train a neural network for detecting urban changes and activities. As pairs, we consider European Remote Sensing (ERS-1/2) and Landsat 5 Thematic Mapper (TM) for 1991–2011 and Sentinel 1 and 2 for 2017–2021. For every era, we use three different urban sites—Limassol, Rotterdam, and Liège—with at least <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>500</mn><mspace width="3.33333pt"></mspace><msup><mrow><mi>km</mi></mrow><mn>2</mn></msup></mrow></semantics></math></inline-formula> each, and deep observation time series with hundreds and up to over a thousand of samples. These sites were selected to represent different challenges in training a common neural network due to atmospheric effects, different geographies, and observation coverage. We train one model for each of the two eras using synthetic but noisy labels, which are created automatically by combining state-of-the-art methods, without the availability of existing ground truth data. To combine the benefit of both remote sensing types, the network models are ensembles of optical- and SAR-specialized sub-networks. We study the sensitivity of urban and impervious changes and the contribution of optical and SAR data to the overall solution. Our implementation and trained models are available publicly to enable others to utilize fully automated continuous urban monitoring.