Spatially Explicit Reconstruction of Cropland Using the Random Forest: A Case Study of the Tuojiang River Basin, China from 1911 to 2010

Abstract

A long-term, high-resolution cropland dataset plays an essential part in accurately and systematically understanding the mechanisms that drive cropland change and its effect on biogeochemical processes. However, current widely used spatially explicit cropland databases are developed according to a simple downscaling model and are associated with low resolution. By combining historical county-level cropland archive data with natural and anthropogenic variables, we developed a random forest model to spatialize the cropland distribution in the Tuojiang River Basin (TRB) during 1911–2010, using a resolution of 30 m. The reconstruction results showed that the cropland in the TRB increased from 1.13 × 104 km2 in 1911 to 1.81 × 104 km2. In comparison with satellite-based data for 1980, the reconstructed dataset approximated the remotely sensed cropland distribution. Our cropland map could capture cropland distribution details better than three widely used public cropland datasets, due to its high spatial heterogeneity and improved spatial resolution. The most critical factors driving the distribution of TRB cropland include nearby-cropland, elevation, and climatic conditions. This newly reconstructed cropland dataset can be used for long-term, accurate regional ecological simulation, and future policymaking. This novel reconstruction approach has the potential to be applied to other land use and cover types via its flexible framework and modifiable parameters.

Keywords

• historical land use
• random forest algorithm
• reconstruction
• spatial distribution