Physical environmental indicators of protected area representativeness: a machine-learning approach across 35 Japanese national parks

Evaluating the physical environmental representativeness of protected area networks requires spatially explicit indicators that are locally calibrated yet nationally comparable. We introduce PhysRep (Physical Representativeness score), a continuous indicator of how closely physical environmental conditions at any location resemble those within a protected area's strictly protected zones. PhysRep is derived from park-specific XGBoost classifiers trained on climatic, topographic, geological and micro-landform variables within an H3 hexagonal grid at approximately 0.1 km 2 resolution, and demonstrated across all 35 Japanese national parks. Spatially grouped nested cross-validation guarded against inflation from spatial autocorrelation, and SHAP values identified the variables driving PhysRep at park and network scales. All 35 park-specific models outperformed label-shuffled negative controls in ROC-AUC (area under the receiver operating characteristic curve, median 0.792 versus 0.488; smallest margin 0.113), the metric used for tier assignment because it is invariant to class prevalence (median F1: 0.777). Climatic and topographic variables dominated nationally, while geological and micro-landform features dominated in regionally specific contexts. Nationwide mapping revealed spatially coherent high-representativeness areas extending beyond current park boundaries, but for most parks these occur as many small, scattered patches rather than large contiguous tracts, indicating limited environmental continuity around individual parks. PhysRep surfaces thus make representativeness gaps and redundancies explicit at the park level. We recommend that protected area network planning under area-based frameworks such as 30 × 30 and Other Effective Area based Conservation Measures (OECMs) incorporate park-specific similarity assessments alongside existing biological and connectivity indicators, prioritizing parks with little contiguous environmental continuity nearby and recognizing where such continuity already provides regional-scale redundancy. The indicator is modular and transferable, requiring only spatial boundaries and environmental rasters as inputs.