Improved Image Classification by Proper Patch Size Selection: TerraSAR-X vs. Sentinel-1A

Abstract

When we perform image content classification by appending semantic labels to regularly cut image patches, we have to be sure that the selected patch size is a good choice for the images at hand. In the following, we look at SAR (Synthetic Aperture Radar) satellite images, and analyse the impact of the selected patch size on the attainable classification accuracy. For test images with precisely known ground truth, one can determine the true precision / recall performance of the applied classification method. In our case, we interactively trained a classifier system via active learning, and compared the resulting classification accuracy for high and medium resolution SAR images of different space borne instruments taken over urban areas, characterized by a high diversity of target categories. At a first glance, it turns out that the selected patch size does have a significant impact leading to a varying number of identified categories with strangely related confidence levels. A fundamental understanding of the relationships between the number of detected categories and their associated confidence levels requires detailed knowledge about SAR imaging, target characteristics, pixel size effects, radiometric image quality, the availability of appropriate semantic labels, and the selected active learning environment together with its image classification tool.