**4. Discussion**

The road detection methodology proposed here aims above all to demonstrate the potential for low cost, but reliable mapping and monitoring of road networks, that can easily be adapted and transferred to extensive regions. It has been applied to three desert areas, each covering around 47,500 km2, each corresponding to the footprint of one Sentinel–1 IW scene. The results over all three areas have achieved an IoU accuracy of over 80%. This accuracy metric takes into account the class imbalance, which is typically the case for road detection in satellite imagery, where in any given area the non–road pixels are expected to greatly outnumber the road pixels. The rank distance is over 75% in all the areas tested, which demonstrates the proximity (in completeness and correctness space) between the reference and model detected roads. The recent study of Abdelfattah and Chokmani [4], who also used Sentinel–1 for road detection over a similar area, could be considered a benchmark to assess the performance of the methodology described here. The correctness and completeness of detections reported in their study are at least 10 percentage points below those described in this paper, for each of the AOIs, and despite the fact that the buffers surrounding the reference objects and detections were larger (three pixels as opposed to two pixels in the study reported here). However, the authors did focus on smaller roads and tracks, while the research of this paper includes major roads in addition to smaller unpaved tracks, so this comparison needs to be treated with caution.

Each desert area in which the methodology was tested has di fferent characteristics, such as diverse sand dune forms, varying predominant road types, and the presence of other infrastructure. Despite these di fferences, the algorithm performs well and therefore demonstrates robustness. Moreover, all the required input is free and open, including the satellite data from which roads are detected, and the OSM reference data. The methodology is hence cost–e ffective. To scale to other areas, the model does require training. Attempts to apply pre–trained models to the other AOIs resulted in a poor performance, probably due in large part to the greatly varying sand dune morphology in each area. However, the U–Net model architecture is particularly e fficient, and training with one NVIDIA GTX1080 Ti GPU never took more than around 25 min. The capability of the algorithm to work well over multiple areas with the available OSM data, without manual intervention for reference dataset cleaning, demonstrates scalability.

Despite the success of the model there are some limitations. While there were few false positives, a significantly higher number of false negatives were encountered, where the algorithm failed to detect roads. This was consistent across all test areas. Many of these false positives were due to the complexity of the context in which the road was situated. Some roads, for example, ran alongside other infrastructure, which in the resolution cell of the SAR input data could be misinterpreted as natural or other non–road features. A possible solution to mitigate these missed detections could be to expand the training set to include more OSM training samples over a wider area, or to include additional classes with a mixed infrastructure.

Another limitation is related to the required input data. In order to reduce speckle while preserving the spatial resolution, speckle filtering was carried out in the temporal domain. This requires processing of a time series, which is computationally expensive, especially for the average coherence generation with SLC format data. However, in most cases it was demonstrated that the VV average backscatter alone produces the best results. Only in one case did the coherence improve the results of the detection, but not significantly. The added value of the coherence is perhaps in those cases where there is a high proportion of roads made from material characterised by a high coherence, i.e., those that have a stable surface, such as paved roads. These would contrast highly with the surrounding sand, where volume decorrelation causes low coherence. In cases where roads are unpaved, or partially sand covered, the coherence is perhaps too noisy, despite the multitemporal averaging, and may degrade the results. Particularly in less developed areas, where there may be fewer paved roads, the coherence processing could therefore be discarded in the interest of a more computationally e fficient algorithm.

The VH backscatter over all sites was much weaker than the VV. As with the coherence, the VH backscatter input only improved the detection results over the Taklimakan Desert site. Here, the VH backscatter was less pronounced over sand dunes, while still high enough over roads to enable their distinction. This may have helped reduce the ambiguity between the high backscatter encountered in the VV polarisation at certain sand dune inclinations with roads. Again, the unique suitability of the VH channel in this area alone may be due to the high relative permittivity of the roads, causing a high enough backscatter even in the weak cross polarisation channel. Elsewhere, however, the low backscatter return of the VH over less reflective roads may have contributed to the degradation of the results.

In terms of the utility of the algorithm for operational road detection in desert areas, the low number of false positives are advantageous in any alert system. Committing resources to detect human activity in remote areas is expensive and time consuming, especially in developing countries such as in North Africa where the means for such activities may be limited. As a monitoring system, the chosen time series length constrains the maximum frequency of monitoring to at least two and a half months. It may be possible to reduce this and still achieve good results. However, any changes significant enough to be observed at the spatial scale of the model are unlikely to occur at temporal frequencies significantly higher than this.

The algorithm proposed here is only a prototype. Improvements could be made for example to reduce the number of missed detections in challenging areas, perhaps by expanding the training set, and including other classes. It could be interesting also to assess the extent to which the Sentinel–1 stripmap (SM) mode may improve detections. While the higher resolution SM mode (around 10 × 10 m in GRD format) is not as systematically available as the Sentinel–1 IW mode, it may nonetheless be useful to detect smaller roads and tracks that may not be resolved in IW mode. This is particularly relevant for security applications [4]. However, the methodology applied with Sentinel–1 IW data nonetheless demonstrates the potential for regular and large–scale mapping and monitoring of desert roads.
