**5. Conclusions**

The methodology proposed here for road detection in desert areas, using Sentinel–1 SAR data as an input and OSM data for training, has the potential to provide a robust, cost–e ffective and scalable solution for the mapping and monitoring of road networks in desert areas. This methodology is still a prototype that has been tested in three areas, each the size of one Sentinel–1 IW scene. More work is required to test its performance over a wider area and over di fferent desert landscape types. Possible improvements with the Sentinel–1 SM mode could be explored. While the accuracy assessments over the AOIs resulted in Jaccard similarity coe fficients above 84% and rank distances of over 75%, more work still needs to be done to improve the accuracy, in particular to reduce the number of missed detections. Future improvements may include the addition of other infrastructure classes, or mixed classes, to account for roads in the proximity of other structures. The methodology may be further tested to quantify model improvement according to the quantity of training data. Additionally, more experimentation can be carried out with additional data augmentation techniques, such as those that modify the intensity of pixels, rather than their spatial position alone. More importantly, the utility of the system needs to be tested by real end users. Its success should be measured against the available systems already in place. Such pre–existing systems are likely to vary between di fferent users and geographic regions. Any improvements should be tailored to meet specific user requirements. The objective of the work presented here is to assess the benefits of EO and open data in combination with deep learning for cost–e ffective and large–scale monitoring. The ambition is to ultimately improve operational road detection and monitoring to support decision–making. With an increasing global population, dynamic migration patterns, and with expanding and evolving road networks, the need for e fficient monitoring systems is ever more critical.

**Author Contributions:** Conceptualization, C.S.; Data curation, C.S., M.L. and A.L.; Formal analysis, C.S.; Investigation, C.S.; Methodology, C.S.; Project administration, C.S. and S.A.; Software, C.S.; Supervision, M.L. and A.L.; Validation, C.S.; Writing—original draft, C.S.; Writing—review & editing, M.L. and A.L. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research received no external funding.

**Acknowledgments:** The authors would like to thank the following for their contribution to this research: Jerin Paul for the modified U–Net model that was adopted in this methodology [36]; the Copernicus programme, which provides free and open access to Sentinel data; Open Street Map, for the free provision of vector data of roads, and Lucio Colaiacomo (SatCen) for the preliminary preparation of this OSM data; CreoDIAS for access to the cloud processing environment for Sentinel–1 data processing; and the Advanced Concepts Team of ESA, for access to Sandy, the GPU sandbox environment.

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