Soil thickness, a critical parameter for hydrological partitioning, ecosystem functioning, and biogeochemical cycling, is challenging to predict spatially in complex karst landscapes-hampered by high heterogeneity, intricate natural/anthropogenic impacts, and rocky desertification. Here, we integrate interpretable machine learning (ML) with rocky desertification information indices (RIs) to enhance soil thickness prediction in typical karst regions. We evaluated six individual ML models and three stacking ensembles (with/without RIs). RIs significantly boosted model explanatory power and consistency (average 7% improvement, 4%-11%), capturing the heterogeneity of soil thickness associated with karst-specific soil degradation processes. Stacking ensembles reduced RMSE (1.33-2.95 cm) and MAE (0.99-2.73 cm); the stacking model with linear regression as meta-model performed best (R2 = 0.47, RMSE = 31.50 cm), while the Cubist base model showed highest accuracy (CCC = 0.63, R2 = 0.45). Shapley additive explanations and permutation feature importance highlighted dominant drivers (rock exposure, vegetation cover, topography), improving transparency. Uncertainty assessments (prediction interval width and prediction interval ratio) validated robustness and identified high-uncertainty areas (steep topography, severe rocky desertification, model disagreement and sparse sampling). Our RIs-integrated model improves soil thickness prediction in karst regions, presents a potentially scalable framework for analogous complex landscapes, advances understanding of soil formation processes in karst systems, and thereby delivers targeted decision support for regional soil management practices.