SC21 Research poster

Urban change detection with remote sensing data covers a wide field of applications like understanding socio-economic impacts, identifying new settlements, or analyzing trends of urban sprawl. It is used for decades. Analyses, however, are usually carried out manually by selecting high-quality samples, restricted to small scale scenarios either temporarily limited or with low spatial or temporal resolution. To process a large amount of available remote sensing observations for a selected period, we propose a fully automated method to train an ensemble of neural networks, without the need to manually select samples. We consider two eras with three sites, each with at least 500 km^2, and deep observation time series with hundreds up to over a thousand combined synthetic aperture radar (SAR) and multispectral optical observations. In order to train such large data sets, we apply data-parallel deep learning with Horovod and multiple NVIDIA Tesla V100 GPUs.

See the poster HERE

Photogallery