MedGm/um6p-gsmi-geoheritage-ml

Research work om machine learning models of physical accessibility for Morocco's geoheritage sites (geosites), # Geosite Accessibility Modeling — Morocco Machine-learning assessment of physical accessibility for Morocco's geoheritage sites (geosites), built at the Geology and Sustainable Mining Institute (GSMI), UM6P. The catalog spans **1,667 geosites** across all eleven administrative regions with labeled sites, of which **939** carry an independently-sourced, citation-traceable accessibility label (*Easy* / *Moderate* / *Difficult*) -- 733 from the original inventory plus 206 added and fully audited in a second labeling pass. Two companion papers cover the work: a national review (`report/geosite_ai_section_2026.pdf`) and a regional comparison (`report/geosite_ai_section_2026_paper2_regional.pdf`). ## Results at a glance | Model | *N* | Validation | Metric | Value | |---|---:|---|---|---:| | Geosite-location favorability | 1,667 | Spatial block CV | AUC | **0.956** | | Guelmim-Oued Noun + Laâyoune (Easy vs. not) | 22 | 500m LOGO-cluster CV | Accuracy | **90.9%** (+31.8pp vs. local baseline) | | Souss-Massa (Difficult vs. not) | 67 | 500m LOGO-cluster CV | Accuracy | 89.6% | | Béni Mellal-Khénifra (Difficult vs. not) | 174 | 500m LOGO-cluster CV | Accuracy | 88.5% | | Eddakhla-Oued Eddahab (Easy vs. not) | 33 | 500m LOGO-cluster CV | Accuracy | 87.9% | | National (Difficult vs. not) | 939 | 500m LOGO-cluster CV | Accuracy | 74.9% | | National (Easy vs. not) | 939 | 500m LOGO-cluster CV | Accuracy | 71.7% | | National (3-class Easy/Moderate/Difficult) | 939 | 500m LOGO-cluster CV | Accuracy | 56.2% | Every model uses the same core terrain/infrastructure feature stack (slope, ruggedness, elevation, distance to highway, distance to settlement, land-cover friction -- regionally, extended with geological-domain or tourism-infrastructure features when that beats the baseline) and a 500m haversine-clustered leave-one-group-out CV protocol throughout, specifically to avoid the near-duplicate-site leakage that inflates naive random-split accuracy in spatial data. National numb …