SAR Image Land Cover Datasets for Classification Benchmarking of Temporal Changes

This is the authors' abstract. We don't add key points for this paper.

The increased availability of high-resolution synthetic aperture radar (SAR) satellite images has led to new civil applications of these data. Among them is the systematic classification of land cover types based on the patterns of settlements or agriculture recorded by SAR imagers, in particular the identification and quantification of temporal changes. A systematic (re)classification shall allow the assignment of continuously updated semantic content labels to local image patches. This necessitates a careful selection of well-defined and discernible categories being contained in the image data that have to be trained and validated. These steps are well-established for optical images, while the peculiar imaging characteristics of SAR sensors often prevent a comparable approach. Especially, the vast range of SAR imaging parameters and the diversity of local targets impact the image product characteristics and need special care. In the following, we present guidelines and practical examples of how to obtain reliable image patch classification results for time series data with a limited number of given training data. We demonstrate that one can avoid the generation of simulated training data if we decompose the classification task into physically meaningful subsets of characteristic target properties and important imaging parameters. Then, the results obtained during training can serve as benchmarking figures for subsequent image classification. This holds for typical remote sensing examples such as coastal monitoring or the characterization of urban areas where we want to understand the transitions between individual land cover categories. For this purpose, a representative dataset can be obtained from the authors. A final proof of our concept is the comparison of classification results of selected target areas obtained by rather different SAR instruments. Despite the instrumental differences, the final results are surprisingly similar.