Detail of mixed infrastructure

**Figure 15.** Close–up of a road segmen<sup>t</sup> in VHR optical data available on Google Earth Pro. The area corresponds with that shown by the red arrow in Figures 12–14. The road segmen<sup>t</sup> was not detected by the model. As evident in the figure, the road runs parallel with other infrastructure, which may have affected the ability of the model to correctly interpret the scene. The imagery date is reported to be 13 September 2013.

#### *3.3. Taklimakan Desert*

Figure 16 shows roads detected by the model over a part of the Taklimakan Desert 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 17. Figure 18 shows the Sentinel–1 input as a red, green, and blue combination of VH and VV backscatter and coherence averages, respectively. Figure 19 shows a Sentinel–2 image of the same area. In this subset there were no OSM roads. The Sentinel–2 image was acquired on 31 July 2019, approximately in the middle of the Sentinel–1 time series (see Table 1).

The confusion matrix for the accuracy assessment is shown in Table 8. Table 9 shows the values of various accuracy indices. The average Jaccard similarity coefficient calculated is 89% and the rank distance is 75%. As with the other areas, there were more false negatives than false positives. Compared to the other areas, there appear to be more paved roads, with straighter paths. The sand dunes are larger and do not have the characteristic lines of high backscatter apparent in the other two areas, although some sparse misclassifications still arise over natural features.

This was the only area where the best results were obtained with all three SAR input layers of coherence, VV backscatter and VH backscatter. The backscatter over sand dunes is much lower in VH than VV, while the road features are still clearly defined, perhaps due to the high relative permittivity of the paved roads. This may be the added value of the VH layer. The coherence layer still displayed much speckle over the sand dunes, but the roads were very clearly defined, perhaps again due to the material of their construction.

Sand drift encroachment on roads in the Taklimakan desert is a serious problem and many efforts have been made to mitigate the issue [7,53,54]. Figure 20 shows a road segmen<sup>t</sup> of the subset in VHR optical data available on Google Earth Pro, the date of which is reported to be 26 October 2014. The road is partially buried in this image, but the model output shows a continuous, unbroken line. It would seem that maintenance had been carried out on this road in between the date of the VHR optical image acquisition and the date range of the Sentinel–1 time series used as an input to the model.

**Figure 16.** Detected roads for part of the Taklimakan Desert 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 17.

Detected roads for randomly selected subset of Taklimakan Desert AOI 

**Figure 17.** Detected roads for a randomly selected 0.2 × 0.2 degree subset over the Taklimakan Desert AOI. White lines correspond with detected roads. The red arrow points to an example of a formerly buried road segment.

Sentinel–1 VH, VV and Coherence input to Taklimakan Desert model 

**Figure 18.** Sentinel–1 colour composite of time series averages of the VH backscatter in red, VV backscatter in green, and coherence in blue. This comprised the input to the model for the Taklimakan Desert area. The extent is the same as that for Figure 17. Contains modified Copernicus Sentinel data 2020.

Sentinel–2 image of Taklimakan Desert AOI subset 

**Figure 19.** Sentinel–2 image of the same area as in Figure 17. The image was acquired on 31 July 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. Contains modified Copernicus Sentinel data 2020.


**Table 8.** Confusion matrix for true and detected roads calculated for the same area as in Figures 17–19.

**Table 9.** Accuracy metrics calculated for the Taklimakan Desert results.

