Are ROC-improved logistic regressions beneficial to statistical models of land change? A counter example

Logistic regression (LR) is widely used in land change modelling; however, its traditional form assumes independent input variables, which is often not realistic. Although the improved models offer better fitting capabilities, it is unclear whether this leads to more accurate land change simulations. To address this gap, we compared the basic LR model with five classic improved models using Lhasa as a case study, comparing the receiver-operating characteristic (ROC) values of each model and further evaluating the performance of the land change models generated by coupling each LR model with the CLUE-s model using six evaluation metrics (Kno, Klocation, Kquantity and the divergence indices (D, A and Q)). The results show that the improved LR models exhibit significantly enhanced ROC values. Specifically, the combined LR achieved the highest average ROC value of 0.941 across different neighbourhood sizes, and the average ROC values of all improved regressions exceeded 0.9, which is significantly higher than that of the ordinary LR (0.872), which remains unaffected by changes in neighbourhood size. However, concerning the land change simulation accuracy, ordinary LR demonstrated a clear advantage, consistently achieving the best performance across all six evaluation metrics regardless of neighbourhood size. Conversely, the improved regressions performed worse, and the combined logistic regression (CL), despite having the highest ROC, performed the poorest in four out of the six evaluation metrics. This finding indicates that there is no inherent link between higher ROC values and improved land change model accuracy. This study further explores the underlying causes of this phenomenon and suggests directions for improvement.