**5. Conclusions**

The main contributions of this work are that we demonstrated the feasibility of vehicle and vessel detection in optical satellite imagery with a variable spatial resolution between 0.3 and 0.5 m, in which these objects are down to a few pixels in size. We selected four single-shot detection network architectures for their fast execution time and compared these, in the meantime optimizing the many hyperparameters for such a case of small object detection in large images. We empirically showed that there is a need to tune the hyperparameters differently for satellite object detection, to reach a good accuracy. A good understanding of the different hyperparameters is primordial for such a task, and thus we explained some of the hyperparameters specific to the loss function of single-shot detection networks. The implementation of these four models can be found in our open-source Lightnet library [22].

From our results, we can conclude that D-YOLO seems to be the most optimal detector, reaching the highest accuracy (APvehicle: 60%, APvessel: 66%) and fastest runtime speeds ( ±4 ms per 416 × 416 patch). While our best results might not be good enough for fully automated detection pipelines, they can already be deployed as a tool for helping data analysts, speeding up their workflow tremendously. It is in this setting that the speed of our networks is primordial, as it promotes a convenient and fast workflow for data analysts and allows them to keep on working with the tool without long waiting times.

The problem of automatic satellite object detection can certainly not be considered solved. There is a dire need for bigger datasets with lots of variability, and even more so for the step of fine-grained classification, where we also need a more balanced dataset, which has multiple examples of each class of vehicle and vessel. The aforementioned technique of using already existing object detection models, such as the ones presented in this paper, but keeping a human in the loop as supervisor could prove to be a valuable tool in order to more easily scale up datasets, after which we might be able to create stronger models, capable of fully automatic small object detection in satellite imagery.

**Author Contributions:** Conceptualization, S.P., V.K. and J.-P.R.; Data curation, T.O., S.P., V.K. and J.-P.R.; Formal analysis, T.O.; Funding acquisition, S.P. and T.G.; Investigation, T.O.; Methodology, T.O. and S.P.; Project administration, S.P., V.K. and T.G.; Software, T.O.; Supervision, S.P. and T.G.; Validation, T.O., S.P., V.K. and J.-P.R.; Visualization, T.O.; Writing—original draft, T.O.; Writing—review and editing, S.P., V.K. and T.G. All authors have read and agreed to the published version of the manuscript.

**Funding:** This work was partially funded by the EU (Tender Lot 6 SatCen RP-01-2018) and FWO (SBO Project Omnidrone).

**Acknowledgments:** We thank the SADL research group of KU Leuven for their project managing efforts during this project and help with annotating.

**Conflicts of Interest:** The authors declare no conflict of interest. The funders provided the raw image data, outlined their needs and current workflow to assess the necessary research and, finally, helped in the writing of this manuscript.

**Disclaimer:** The views and opinions expressed in this article are those of the authors solely and do not reflect any official policy or position of their employers.
