**4. Conclusions**

This study investigated the detection of plant diseases using hyperspectral cameras at ground and UAV scales. In this context, the appropriate data analysis was decisively able to reach suitable results. Supervised classification has the advantage of separating disease-related signals from a huge amount of natural biological, geometrical, and sensor-related variability within a hyperspectral image of a crop canopy in the field. We proved that hyperspectral imaging in combination with supervised classification and regression showed good accordance to visual assessment at the ground. This allows questions to be addressed regarding the transfer of information between different scales and sensors. We showed that a feature selection was able to increase the prediction accuracy if it was performed on the analyzed data set. In contrast, scale or sensor transfer of selected feature sets was not successful, and was even less predictive than an uninformed regularly sampled feature set. This highlighted the importance of a precise specification of a prediction task by representative data samples. Deviations in data characteristics can significantly impair the performance of a data analysis pipeline or a tailored sensor in real-life applications.

This study sets a basis for ongoing research. New, upcoming sensors fulfilling the demands defined in this study might also cope with the current disadvantages. Consequently, there is a high probability that the defined flying sensor system with high resolution spectral camera, computing capabilities, and self-localization will be realized. Adapted legal conditions would allow an integrated system of field managing software, remote sensing based predictions, and current observations from the field using an automatized UAV.

**Author Contributions:** Conceptualization, D.B., J.B. and A.-K.M.; methodology, D.B. and J.B.; software, D.B. and J.B.; validation, D.B. and J.B.; formal analysis, D.B. and J.B.; investigation, D.B., J.B. and A.-K.M.; resources, D.B., J.B. and A.-K.M.; data curation, D.B., J.B. and A.-K.M.; writing—original draft preparation, D.B. and J.B.; writing—review and editing, D.B., J.B. and A.-K.M.; visualization, D.B., J.B. and A.-K.M.; supervision, J.B. and A.-K.M.; project administration, A.-K.M.; funding acquisition, A.-K.M.

**Funding:** This work was funded by BASF Digital Farming.

**Acknowledgments:** The authors would like to thank Onno Muller (Research Center Jülich, Germany) for providing the basic Phytobike frame, Thorsten Kraska for general support at Campus Klein-Altendorf and Winfried Bungert (Campus Klein-Altendorf, Germany) for implementation of cultivation measures during the vegetation period.

**Conflicts of Interest:** The authors declare no conflict of interest.
