Identification and apportionment of the drivers of land use change on a regional scale: Unbiased recursive partitioning-based stochastic model application

Abstract

Land-system science places increasing emphasis on the complex, telecoupled and often nonlinear nature of land-system changes. Categorical diversity, spatial-temporal heterogeneity and the rapid change rate of land use drivers challenged researchers to determine this sophisticated nature of landscape changes at the regional scale. This study aims to develop a new framework, involving remote sensing, GIS and machine learning, to identify and apportion the important factors responsible for land use change at the regional scale. The Jiangxi province in China was used as a case study. The drivers of land use change were identified and apportioned using stochastic models based on unbiased recursive partitioning method embracing the conditional inference tree (CIT) and random forest (RF) with a focus on cropland and urban land. Regression trees for determining the major drivers of cropland and urban land change were established. Partial dependence plots and two-variable interaction plots from the resulting RF models for explaining the contributions of the drivers of cropland and urban land change were developed. A spatial autoregressive model was implemented as a supplement tool to help explain the causes of land use change. The determinants of cropland and urban land change were quantitatively assessed by CIT and RF along with the interactions of multiple drivers. The models were verified using rigorous out-of-sample and actual-versus-predicted testing. The results show strong suitability of unbiased recursive partitioning-based models to the assessment of the complex drivers of land use change at a regional scale