The Use of Cascaded Learning for TerraSAR-X Image Classification

The abundance of available satellite images calls for their automated analysis and interpretation, includ-ing the semantic annotation of discovered objects as well as the monitoring of changes within image time series. A common approach is to cut large satellite image into contiguous patches and to classify each patch separately by attaching a semantic patch content label to it. In this context, the selected patch size is a critical parameter, as patches being too large may contain multiple objects and patches being too small may not be understandable due to missing contextual information. Therefore, we advo-cate a “cascaded” strategy where, when necessary, large patches are iteratively decomposed into small-er sub-patches until a clear semantic content understanding has been reached. This strategy can be em-bedded into an interactive active learning and exploitation environment where high classification effi-ciency can be reached by skipping unnecessary decomposition steps. The resulting local multi-level off-spring statistics is indicative of the recorded land cover category. In the following, we report about our experiences with medium and high resolution Synthetic Aperture Radar (SAR) image classification when using such a cascaded learning approach. The most important phenomenon is the impact of image reso-lution. The higher the resolution, the higher the number of discernible land cover categories, in particu-lar for built-up areas and industrial sites where we can see and interpret the impact of distinct human-made activities. Here, the offspring statistics depends on the actual image resolution. This becomes ap-parent when we compare the same target areas acquired by different space-borne SAR sensors (e.g., Sentinel-1A versus TerraSAR-X). In addition, it turns out that several country-specific regional surface cover categories can be trained and retrieved with SAR images that often appear differently in optical satellite images; however, any increase in classification accuracy has to be paid for by higher computa-tional effort