A SAR Benchmarking Tool for Generation of Image Datasets: Applications for Urban Areas

Kurzfassung

During the last years, we saw a growing interest in satellite image analysis including semantic and quantitative content description of images, physics-related classification of image segments, the quest for the causes and effects of changes over short and long time periods, and disaster relief support tools. These remote sensing applications are most often using optical datasets; fewer users employ SAR (Synthetic Aperture Radar) datasets. However, both cases call for the identification of land cover / land use details within the full image area. As modern imaging sensors are characterized by high spatial resolution and a wide field of view, we face a high diversity of target types, target objects, as well as temporal changes and their spatial interrelationships. Therefore, we need reliable benchmarking tools to ascertain the actual quality of our image analysis and retrieval results. Primary benchmarking tools for automated benchmarking are collections of selected reference images with known semantic content and characteristics that can be employed for quantitative and comparative image analyses. For optical sensors there exist several well-known and publicly available datasets comprising typical remote sensing image patches, while comparable SAR datasets are very scarce. For our user-oriented application cases no well-known high-resolution and publicly available SAR reference datasets exist. In our case, we collected from all over the world 1,000 urban and industrial areas together with their infrastructure TerraSAR-X images and 75 Sentinel-1 images [1, 2], and assigned them to typical cases of acquisition parameters and target areas. Then about 30% of these images were selected for training and generating a reference dataset using our benchmarking data mining system. We tiled these training images into about 250,000 patches (in case of TerraSAR-X) and 180,000 patches (in case of Sentinel-1), and labeled each patch with a semantic category. The selection of the training images was controlled by SAR experts and was driven by the typical requirements of our application cases and their diversity. When we aim at a systematic and universal benchmarking of our classification quality, we need an established procedure that includes the methodology described in [3], complemented by a manual annotation for the remaining patches left unclassified by the data mining system, and a visual inspection of the results. The basic finding of this paper is the verification that we can perform and obtain an automated generation of benchmarking SAR datasets with good quantitative performance measures. [1] TerraSAR-X archive portal, 2018. Available: http://eoweb.dlr.de/. [2] Sentinels Scientific Data Hub, 2018. Available: https://scihub.copernicus.eu/dhus/#/home. [3] C. O. Dumitru, G. Schwarz, and M. Datcu, “Land Cover Semantic Annotation Derived from High-Resolution SAR Images”, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 9(6), pp. 2215-2232, 2016.