Figure 1. Location of the study area

Data collection

We used Sentinel 2A and Landsat 5–8 series satellite data downloaded by the United States Geological Survey USGS (https://www.usgs.gov/) to draw the 2010–2020 land use type map, divided into five categories: cropland, building, forestland, grassland, and water. Vegetation information was also extracted to obtain the time series of vegetation coverage in the Jiuyuangou watershed from 2010 to 2020. The 12.5 m resolution DEM data was downloaded through NASA (https://search.asf.alaska.edu/#/) to extract topographic parameters of the watershed and the slope length and slope factor. We downloaded the daily rainfall dataset from 1960 to 2020 from the National Meteorological Data Center (https://data.cma.cn/) and performed a spatial interpolation of meteorological data based on this data to obtain the 2010–2020 rainfall erosivity layer of the basin. The soil physicochemical properties of the watershed were obtained from the FAO 250 m resolution global soil texture survey data (https://soilgrids.org/). We used Google high-resolution historical imagery to identify soil and water conservation tillage measures in the watershed from 2010 to 2020.

Model implementation

Liu et al. (2002) innovatively proposed the vegetation measure (B factor), engineering measure (E factor), and tillage measure (T factor) according to the actual soil and water conservation situation on the Loess Plateau and combined them with the RUSLE equation to establish a Chinese soil erosion prediction model suitable for the Loess Plateau:
where A is the soil erosion modulus based on the CSLE model (1 t ha–1 yr–1=100 t km–2 yr–1), R is rainfall erosivity (MJ∙mm·ha–1∙h–1∙yr–1), K is soil erodibility factor (t∙ha∙h–1·ha–1∙MJ–1∙mm–1), L is slope length factor, S is slope factor, B is vegetation coverage measure factor, E is water and soil conservation engineering measure factor, and T is water and soil conservation tillage measure factor.
We obtained the rainfall erosivity grid layer using the rainfall erosivity model based on daily rainfall data from meteorological stations around the Jiuyuangou watershed. For soil erodibility factor (K factor) in the Loess Plateau region, most studies have adopted the soil erosion and productivity impact estimation model (EPIC; Williams et al., 1983), which uses soil organic matter and particle composition for estimation. Based on 12.5m ALOS DEM data (https://www.earthdata.nasa.gov/), the LS_TOOL slope length and slope factor calculation software developed by Zhang et al. (2017) was used to complete the terrain factor calculations, and ArcGIS 10.7 was used to generate a raster layer of slope length and slope factors. In this study, based on the ‘Technical Regulations for Dynamic Monitoring of Regional Soil Erosion’ from the Department of Soil and Water Conservation, Ministry of Water Resources, China, the vegetation coverage measure (B factor) was calculated using the vegetation coverage combined with land use type data and monthly rainfall erosion ratio with month as the time step. The E factor of the watershed was obtained by interpreting Google historical imagery within the study area. The raster calculator in ArcGIS was used to assign the T factor to the corresponding slope grading map and obtain the T factor map of soil and water conservation tillage measures.

Land cover and land use changes and soil erosion changes (2010–2020)

For soil erosion change and LULCC assessment, we used the transition matrix method, where diagonal values show the 2010–2020 stable area of LULC and soil erosion grades. The soil erosion modulus map and LULC map combined with the transition matrix help us understand the spatial and temporal evolution of land use patterns and soil erosion grades in the Jiuyuangou watershed. As per Gilani et al (2021), we stipulate soil erosion transformations from lower to higher erosion grades as ‘loss’ and higher to lower erosion grades as ‘gain’. For changes in the spatial pattern of land use, we specify the conversion of other land use types to cropland as ‘loss’ and cropland to other land use types as ‘gain’. On this basis, we conduct a binary variable spatial correlation analysis to determine how the two variables change simultaneously in the whole watershed space (Nandi and Shakoor, 2010). Finally, we produced a bivariate choropleth to understand the spatial pattern of soil erosion and LULCC and the association of gain, loss, and no change between 2010 and 2020.