**3. Results**

The model performed well in all areas, despite the diverse landscape forms and road conditions encountered in each. Especially in VV polarisation, many sand dunes produced a high backscatter. This is typical when the incidence angle of the SAR system equals the angle of repose of sand dunes [52]. The model proved nonetheless capable of distinguishing roads from sand dunes, rock formations and other natural features with similar backscatter characteristics as roads.

Of the various SAR input data types, the VV backscatter average alone proved the most e ffective for both the North Sinai Desert and Grand Erg Oriental. Only for the Taklimakan Desert site did all three layers of coherence, VV backscatter and VH backscatter yield the best results. Table 3 lists the IoU accuracies for each of the single SAR input layers, and for all three input layers for each of the AOIs. The fact that the use of all the input layers improved the results in only one area shows that more information provided to a model is not necessarily better. The decreased accuracy caused by the inclusion of the VH backscatter and coherence is perhaps due to the increase in speckle in these layers. This may have hindered the models in distinguishing particularly challenging roads, such as those that may be narrow, unpaved, or partially buried. The VV backscatter return over this type

of landscape is generally much stronger than the VH backscatter, and enables a clearer distinction of roads. The exception of the Taklimakan desert is perhaps due to the predominant type of road surface and surrounding context of sand dunes. The results for each area are discussed in more detail in the subsections below.


**Table 3.** Intersection over union (IoU) accuracy of road detection models with di fferent SAR input types. The best results for each area are highlighted in bold.

#### *3.1. North Sinai Desert*

Figure 6 shows roads detected by the model over a part of the North Sinai AOI. The area includes the location of the randomly selected subset (0.2 × 0.2 degree area) in which a more accurate performance evaluation was carried out. This subset is shown in more detail in Figure 7. Figure 8 shows the corresponding area of the SAR layer used as a model input, which was the Sentinel–1 multitemporal backscatter average of the VV polarisation only. Figure 9 shows a Sentinel–2 true colour image of the subset with the available OSM data for this area overlaid. The Sentinel–2 image was acquired on 2 August 2019, which is approximately in the middle of the Sentinel–1 time series (see Table 1).

The confusion matrix for the accuracy assessment is shown in Table 4. Table 5 shows the values of various accuracy indices. The average Jaccard similarity coe fficient is 89% and the rank distance is 80%. There were few false positives, i.e., non–roads classified as roads, despite the many natural features of high backscatter that could have been misinterpreted by the model as roads, such as sand dune ridges. However, there were many more false negatives, i.e., undetected road segments. In some cases, the broken segments shown in the mask were correct in that the actual road was partially buried in segments (see Figure 10 and arrows in Figures 7–9).

The VH backscatter over the entire area was much lower, and road features much less clearly defined, particularly those that were unpaved and partially sand covered. This may be the reason why the VH backscatter input degraded the results. The coherence layer highlighted very clearly the roads, but overall had much more speckle, which is possibly why this also reduced the quality of the detection.

**Table 4.** Confusion matrix for true and detected roads, with only the VV backscatter as input, calculated for the same area as in Figures 7–9.



**Figure 6.** Detected roads for part of the North Sinai AOI. The yellow rectangle shows a 0.2 × 0.2 degree subset over which roads were manually digitised and a performance evaluation carried out. This area is shown in more detail in Figure 7.

Detected roads for randomly selected subset of North Sinai AOI 

**Figure 7.** Detected roads for a randomly selected 0.2 × 0.2 degree subset over the North Sinai area. White lines correspond with detected roads. The red arrow points to an example of a buried road segment.

Sentinel–1 VV backscsatter average input to North Sinai model

**Figure 8.** The Sentinel–1 average vertical transmit–vertical receive (VV) backscatter used as an input to the model for North Sinai. The area is the same as that of Figure 7. Contains modified Copernicus Sentinel–1 data 2020.

**Figure 9.** Sentinel–2 image of the same area as in Figure 7. The image was acquired on 2 August 2019 (roughly in the middle of the Sentinel–1 time series). It is displayed in true colour, bands 4,3,2 as red,

green, and blue, respectively. Overlaid in red are the available OSM roads for this area. Contains modified Copernicus Sentinel–2 data 2020.
