On the Influence of Sampling Strategies for Classification of Remote Sensing Data

Methods for object detection from optical near-range photographs often have access to databases containing thousands or even more labelled images. These images are used to train machine-learning based approaches and to evaluate their performance. In contrast, remotely sensed data is often more difficult to obtain and to be labelled, in particular for sensors such as synthetic aperture radar (SAR) or hyperspectral cameras. Many methods proposed in the literature are trained and evaluated on a single image. Training and testing data are still ensured to be disjoint. Nevertheless, the particular sampling strategy to create these datasets from a single image has an influence on the final classification performance.