The Dongzhi tableland (34°50′-37°19′N, 106°14′-108°42′E) lies in Qingyang city, Gansu province, China, with an area of 910 km² and a slope of 0°-5° in the southern part of the CLP (Figure 1). The loess thickness in the Dongzhi tableland ranges from 100 m to 300 m (Zhu et al., 2018) and consists of aeolian dust deposition during the Quaternary era (Edward et al., 1995). The average elevation of the Dongzhi tableland is approximately 1450 m and inclines from the northeast to the southwest. The area has a typical Forman continental monsoon climate, with 80%-90% of annual precipitation occurring between June and September. The annual precipitation is 450-600 mm, and the annual evaporation is 1000-3000 mm. The slopes ranging from 0° to 5° are concentrated in the center of the tableland (Figure 1b). The Dongzhi tableland is now facing the threat of severe gully erosion due to rapid urbanization.

Road maps of the Dongzhi tableland in 2005 and 2020 showed that most of the roads were distributed on flat plateau surfaces with slopes of 0°-5°. The road density increased significantly from 2005 to 2020 (Figure 3). In 2005, the total road length on the Dongzhi tableland was approximately 3095 km, with a density of 1.09 km/km². The county-level roads showed the highest density (0.93 km/km²) among the five types of roads. In 2020, the total length of roads on the Dongzhi tableland was approximately 4274 km, with a density of 1.51 km/km². The county-level roads also showed the highest density among the five types of roads with a total of 1.24 km/km². During the past 15 years, the total length of roads on the Dongzhi tableland increased by 38.1%, and the annual growth rates of the five types of roads were 7.29 km/a for expressways, 0 km/a for national roads, 5.95 km/a for provincial roads, 6.45 km/a for urban-level roads and 58.86 km/a for county-level roads (Table 2).

For the density mapping, the KDE results demonstrated that the road kernel density of the Dongzhi tableland significantly increased from 2005 to 2020. The geometrical shape of the kernel density distribution showed a concentrated line, with the highest density spread occurring along the flat plateau surface of the tableland (Figure 4). In 2005, the downtown areas of Ningxian, Xifeng, and Qingcheng counties showed the highest kernel density. In 2020, the regions with a high kernel density of roads greatly expanded and spread around the former areas, indicating that the road network on the Dongzhi tableland had rapidly expanded during the past 15 years. Xifeng district exhibited fast road kernel development, while Qingcheng and Heshui counties showed only a small enhancement in the road network in their suburban regions. In 2005, the low-density road network accounted for 54.28% of the total area, and the extremely high-density road network accounted for only 1.98% of the total area. In 2020, the area ratio of medium-density areas showed the largest proportion, accounting for 44.96% of the total area, and the area ratio of extremely high-density areas increased to 7.88% (Table 3).

The observed land use changes showed that the main land use types on the Dongzhi tableland were grassland and cultivated land in 2005 and 2020, and the area of constructed land had increased greatly by 82.24 km² over the 15-year period (Figure 5). Land use transformation analysis showed that 5.34% of the cultivated land and 6.97% of the grassland had been converted to construction land from 2005 to 2020 (Figure 6). According to the change in landscape indices on the Dongzhi tableland from 2005 to 2020, the landscape indices of LPI, PD, and AI decreased from 28.28, 7.06, and 90.6 to 21.55, 6.95, and 90.4, respectively (Table 4). These ecological factors decreased from 2005 to 2020 in almost all the counties, except the PD value in Xifeng district, which increased by 2.47. This indicated that the landscape pattern of Xifeng district had been continuously segmented and broken and showed strong landscape heterogeneity in spatial form. Meanwhile, landscape analysis demonstrated that different land use types had different landscape metrics (Table 5). For the land use type of cultivated land, the area decreased by 2.64%, and the landscape metrics of LPI, PD, and AI synchronously decreased by 6.73, 0.05, and 0.5, respectively. Contrary to the cultivated land area, the area of constructed land increased by 2.9%, which resulted in the landscape metrics of LPI, PD, and AI increasing by 0.69, 0.06, and 0.27, respectively. However, there was no significant change in the landscape metrics of other land use types between 2005 and 2020.

The area proportion of different land use types exhibiting different degrees of road kernel density in 2005 and 2020 showed that the low- and medium-density road network classes were mostly distributed in grassland, while the high-density road network class was mainly distributed in cultivated land (Figure 7). Over the past 15 years, the proportion of constructed land characterized by the extremely high-density class significantly increased in area by 21.85%. The spatial variations in road kernel density and landscape metrics showed that the landscape patterns on the Dongzhi tableland were greatly affected by the expansion of the road network. The results demonstrated that the road network kernel density was positively and significantly correlated with LPI and AI (p < 0.01) but negatively correlated with PD (p < 0.01). Meanwhile, the correlation coefficients between road density and landscape metrics increased from 2005 to 2020 (Table 6).

The Dongzhi tableland is the main agricultural area on the CLP. In the area, cultivated land is the primary type of land use (Figure 5). Taking the edge line of the tableland as the boundary, the cultivated land and constructed land were mainly distributed across the flat terrain, with slopes ranging from 0° to 5°, while the grassland and woodland areas were mostly distributed in areas with slopes greater than 5°. Due to rapid urbanization, the area of cultivated land has decreased by 2.64% over the past 15 years on the Dongzhi tableland. The landscape metrics of the cultivated land, e.g., LPI, PD, and AI, also decreased, which indicated that the cultivated land had become more fragmented. As the cultivated land was mainly converted to constructed land, the aggregation degree of constructed land and the associated landscape metrics (LPI, PD, and AI) increased.

For the road kernel density analysis, old kernels were mainly distributed in Xifeng district and Ningxian county in 2005. New kernels extended along the old kernel in Xifeng, and a new kernel occurred in Ningxian. Over the past 15 years, the number of main kernels increased from 2 to 3. Meanwhile, the total area of the main kernels on the Dongzhi tableland increased from 55.73 km² in 2005 to 223.55 km² in 2020, representing a 4-fold increase. In Xifeng, there is one main kernel, and that number did not change between 2005 and 2020, while the kernel area greatly expanded from 3.66 to 134.83 km², representing a 36.9-fold increase (Figure 8). Moreover, road networks were mostly distributed throughout areas with slopes between 0° and 5°, while the low and medium densities of road networks were mostly distributed in grassland areas with slopes greater than 5°. This phenomenon is similar to previous findings. Fischer (2018) found that road networks with high density were generally distributed in plain areas with a large population, while the road network in steep terrain tended to be branched and had low density. In this study, it was determined that road network expansion has a close relationship with urban development. For example, the proportion of constructed land with an extremely high density of road networks had increased by 21.85% over the past 15 years (Figure 9). Moreover, Nyssen et al. (2002) noted that road developers generally avoid toe slopes to reduce the cost of bridge construction, but there were implicit costs accompanied by gullying in slope land and pediments, such as soil erosion.

As a precedent and superseding infrastructural requirement for urban growth, road construction would result in substantial landscape changes (Liu et al., 2008). Mann et al. (2021) demonstrated that there were significant relationships between road networks, landscape ecological risks, and topographic factors. In this study, we found similar results to those of Mann et al. (2021). On the Dongzhi tableland, the kernel density and landscape metric changes were concentrated in the flat area from 2005 to 2020, and the road kernel density exhibited significant relationships with the landscape metrics (Figure 9).

With the expansion of the road network, we found that the tableland area was rapidly shrinking. The Dongzhi tableland was reduced by a total of 83.14 km² between 1972 and 2020. In 1972-2005, the average rate of tableland retreat was 0.92 km²/a, and the retreat rate increased to 3.5 km²/a from 2005 to 2020 (Jin et al., 2023). In our previous study, we found that there was a significantly positive relationship between gully expansion and urban area development (Yang et al., 2020). Although the study only characterized a typical gully near the urban area of the Xifeng on the Dongzhi tableland, it may reflect that urbanization has a strong impact on gully expansion and tableland retreat. In this study, we can conclude that such significant tableland retreat is probably related to the rapid expansion of road networks and urban construction in the area, which can decrease rainfall infiltration, increase surface runoff and consequently exacerbate headward erosion and gully expansion. Furthermore, cultivated and constructed land had a stronger impact on road kernel density than other landscape types because the high-density road network was mostly distributed in the cultivated and constructed lands. The two land use types exhibited high degrees of aggregation. Thus, the road kernel density was positively and significantly correlated with the landscape index of the AI (p < 0.01). The urban areas are generally concentrated in the tableland, and their fragmentation degrees are low because they have fewer patches. In contrast, forestland and grassland are located far away from urban areas and are easily affected by terrain conditions (Mo et al., 2017). Therefore, the fragmentation degrees of forestland and grassland are high because they have more patches. Although road network expansion is often reported to induce landscape fragmentation, road networks could also allow construction land to continually expand (Cai et al., 2013). Thus, although the number of construction land patches is small, its area is large, which consequently leads to a relationship between PD enhancement and road kernel density reduction.

Road construction can permanently alter the geomorphic and hydrological settings of the landscape and elevate the potential risk of soil erosion. Previous studies have demonstrated that road construction is an important factor in accelerating gully erosion (Ziegler and Giambelluca, 1997; Kakembo, 2000; Zhang et al., 2009). Nyssen et al. (2002) showed that a 6.5 km long road segment had formed 16 new gullies and caused five small gullies to become inactive. Moreover, the road kernel density was negatively correlated with the gully kernel density. For a clearer interpretation, the gully distribution along a national road of 66.31 km was analyzed on the Dongzhi tableland (Figure 10a). A buffer analysis was conducted, and the threshold value was set to 500 m. We found that 130 gullies were located in the buffering area and that the gully density was 1.02 km/km². In the buffering area, the gully distribution is stable, and the degree of crushing is severe (Figure 10b). Although remote sensing images cannot identify the area change in specific gullies between 2005 and 2020 on the Dongzhi tableland, the area of the Dongzhi tableland decreased from 990.63 km² in 1972 to 960.08 km² in 2005 due to gully erosion. With the acceleration of headward erosion and bank collapse, the tableland further retreated, and the area decreased to 907.49 km² in 2020. In total, the area of the Dongzhi tableland shrank by 83.14 km² between 1972 and 2020, which was 30.55 km² in 1972-2005 and 52.59 km² in 2005-2020 (Jin et al., 2023). Through an investigation of soil erosion control practices on the Dongzhi tableland, it can be observed that the use of culverts is the main drainage method; these culverts receive runoff along the road and channels the water into the gully, which could greatly increase the speed of gravitational erosion and cause landslides (Figure 10c). Thus, appropriate facility designs should be established to avoid new gully formation and development during road design and building processes. At present, there are four kinds of gully consolidation and tableland protection modes on the CLP: (1) simple backfill of the gully head; (2) backfill of the gully head associated with drainage facilities; (3) backfill of the gully head combined with drainage facilities and slope protection; and (4) backfill of the gully head combined with drainage facilities and ecological slope protection, among which the fourth mode is the most effective in tableland protection and gully erosion control (Wang et al., 2019).

This study is valuable for road network construction and ecosystem management, and the kernel density method can provide a basis for exploring the relationship between road network expansion and landscape change. However, the complexity of road network expansion and its impacts on regional landscape patterns and risks are diverse (Barandica et al., 2014; Staab et al., 2015). Studies on road network structure include either a single main road or a complex road network. Thus, more methods should be explored to assess the complex relationships between roads and landscape ecology (Gong et al., 2021). Many studies have demonstrated that unpaved rural roads can facilitate hydrological connectivity, pollutant mobility, and sediment transport in the drainage network and may impact water quality in streams and reservoirs (Mamede et al., 2018), as well as damage ecosystem health (Wemple et al., 2018). Meanwhile, road construction, building activity, and unreasonable land use accelerate soil erosion and consequently lead to land degradation (Song et al., 2018; Kuklina et al., 2021; Mauri et al., 2022). Thus, it is necessary to analyze the relationships between unpaved rural roads and soil erosion sensitivity on the Dongzhi tableland in the future. Moreover, since the late 1980s, the Dongzhi tableland has experienced dramatic land use and land cover changes due to rapid urbanization and ecological project initiation (Cai et al., 2019; Song et al., 2020). This study only analyzed the road network and land use change between 2005 and 2020 due to limited source data. More time samples should be used to study the road network evolution and regional ecological risk in the tableland area.