Exploring the non-linear relations between the cropland expansion and driving factors in China

Exploring the relationships between cropland expansion and their driving factors is crucial to implement cropland protection policies. Previous research primarily explored their linear relationships but seldom discussed the non-linear effects that better reflect the changes in the real world. Thus, a machine learning method, namely, the gradient boosting decision tree (GBDT) model, is introduced to identify the threshold effect and the non-linear correlations between cropland expansion and driving factors in the middle reaches of the Yangtze River urban agglomerations (MRYR), China. Results show the following: (1) During 2000-2020, a 13,141.76 km2 of cropland is expended by taking in other land use types, among which forest and construction land have the top proportions. (2) The relative importance (RI) identified by the GBDT model indicated that area of the county (AC) and cropland area (CA) were the most important variables that affected cropland expansion values of RI at 56.94% and 12.14%, respectively. (3) A non-linear correlation exists between cropland expansion and each explanatory variable, which can be represented by a curve with thresholds. In addition, three correlation types were identified: positive, negative, and piecewise non-linear correlations. This research emphasizes that the non-linear relationships and threshold effect of the socioeconomic development factors on cropland protection should be considered when formulating cropland protection policies.