Despite algorithmic advances in the field of image processing as well as modern computer architectures, automated sensor data processing that guarantees reliable and robust information retrieval in dynamic environmental and varying flight conditions is still a challenging task within unmanned surveillance and reconnaissance missions. In our paper we will elaborate the reasons and propose a promising way out by adapting to variable environmental conditions and states of the UAS platform in terms of the dedicated usage of specialized sensor data processing chains.

However, these specialized chains must be used within their operation space. Otherwise, their performance in terms of detection precision and recall will degrade. To overcome this drawback, we propose to apply chain performance models based on Bayesian Networks (BNs). The evaluation of the BNs takes place during the flight depending on environmental influences. Accordingly, a performance probability can be predicted for each chain, which is used for an automatic chain selection.

We validate our approach within a real flight Search and Rescue scenario (SAR). To compare generalized and specialized chains, we conducted several flight experiments with an EO/IR mission sensor setup: a) to annotated and preselect/filter training-datasets, used for transfer-learning of the Convolutional Neural Networks (CNNs), and b) to derive test datasets in different environmental situations and with varying platform/sensor states. The sensor data comprises variations in illumination and meteorological conditions, photographic conditions (e.g., different sensor elevation angles and ground sample distances) as well as topographic conditions. We provide a comprehensive insight into the detection results derived. Based on these results we conclude that a targeted use of specialized CNNs can outperform generalized CNNs.