Figure 1

Location of meteorological stations and geographical division in Shaanxi Province, China"

Table 1

The datasets sources"

Figure 2

Spatial patterns of SC (a), change rate of SC (b) and the significant (P<0.05) change areas (c) from 2000 to 2013. The change rate at cell level was calculated by using the least square method (LSM)."

Figure 3

Hotspots and coldspots with different confidence levels (The double star (**) and single star (*) superscript indicate hotspots or coldspots are significant at 99% and 95% level respectively)"

Table 2

Statistics on the hotspots and coldspots of soil conservation service in Shaanxi Province, China"

Figure 4

LULC types in hotspots and coldspots. The pie charts indicate the percentage of each LULC from the total area of hotspots or coldspots."

Figure 5

Comparison of LULC change between 2000 and 2013: (a) the area (AR) of LULC change and the mean SC provided by each LULC type; (b) the distribution of other types of LULC transferred to woodland and grassland"

Figure 6

Spatial patterns of mean annual precipitation (PPT) (a), temperature (Tem) (c) and their change rates (b and d) from 2000 to 2013. The change rate at grid cell level was calculated by using the least square method (LSM)."

Figure 7

Comparison of three hotspot analysis methods: KDE (1), Local Moran’s I (2), Getis-Ord Gi* statistics (3)Note: When input features are spatially scattered (see Figure 7a), the KDE can only identify spatial cluster, but it mixes the cluster of high values (i.e., hotspots) and the cluster of low values (coldspots); that is, the KDE can neither tell what the cluster is nor whether it is significant. Fortunately, the Local Moran’s I and Getis-Ord Gi* statistics can both make it. The difference is that the Local Moran’s I is more efficient at identifying the outliers (see Figure 7a-(2) above), while the Gi* statistics is even better at identifying statistically significant hotspots and coldspots with different confidence levels (see Figure 7a-(3)). When input features are spatially uniform grid data (see Figure 7b), the KDE becomes inefficient, while the Local Moran’s I and Gi* statistics work well in this case, and the Gi* statistics especially holds its unique superiority."