Comparative performance of maximum entropy, frequency ratio, and logistic regression models for landslide susceptibility mapping in Semarang Regency, Indonesia

Landslide susceptibility mapping in many regions must rely on inventories compiled from community reporting, which provide presence data only and no field-verified non-occurrence points.Which modelling approach best suits such data remains unclear.This study compares Maximum Entropy (MaxEnt), Frequency Ratio (FR), and logistic regression (LR) for landslide susceptibility mapping in Semarang Regency, Indonesia, using an identical set of conditioning factors and an identical validation scheme so that differences in performance reflect the methods themselves.An inventory of 445 landslide points, compiled from the records of the Regional Disaster Management Agency for 2021-2025 and from field survey, was modelled against 13 continuous conditioning factors at 30 m resolution and evaluated through spatial-block cross-validation (5 km blocks, five folds, ten random block assignments), with random five-fold cross-validation as a comparison.Discrimination was assessed with AUC, TSS at a threshold fixed on the training data, and McClish-standardised partial AUC restricted to the high-specificity range (0.8-1.0), with differences tested using the DeLong test and a paired bootstrap.MaxEnt and FR performed equivalently on the full AUC (0.688 and 0.691, mean of the ten block assignments; DeLong p < 0.05 in one of ten) and on TSS (0.278 and 0.269), both within the range of moderate discrimination.LR scored lower (AUC 0.653; TSS 0.222) and differed significantly from both, a gap that persisted across 1,000 pseudo-absence datasets and when pseudo-absences were kept at least 500 m or 1000 m from landslide points.MaxEnt outperformed FR only over the high-specificity range (pAUC 0.610 against 0.578; bootstrap p < 0.005 in all ten block assignments) and in the most susceptible class, where it captured 57.5% of events against 52.4% for FR and 45.6% for LR.The ranking of the models was unchanged when the inventory was thinned to a minimum of 100 m or 250 m between points.Slope was the strongest predictor in both models that produce importance scores.Presence-based modelling is therefore adequate for report-based inventories, with FR suited to rapid and transparent mapping and MaxEnt to sharpening the most susceptible zone.