Accurate maps of alpine peatlands are essential for conservation and restoration planning, yet existing inventories remain incomplete, particularly for small and spatially isolated sites. Remote sensing improves peatland detection; however, robust classification requires sampling strategies that account for the spatial structure of reference data. We developed a multi-source Random Forest framework to map acid and base-rich peatlands habitats in two environmentally contrasting Italian Alpine basins: Orco in the Western Alps and Avisio in the Eastern Alps. The models integrated Sentinel-1 radar backscatter, Sentinel-2 multispectral bands and optical indices, and hydro-geomorphological variables derived from the Copernicus GLO-30 digital surface model. We compared three classification strategies representing different reference data conditions and mapping objectives: a random point-based three-class classification using spatially distributed reference data within each basin, a cluster-based three-class classification simulating spatially incomplete reference data, and a cluster-based binary classification testing peatland detection under the same clustered-data conditions. The random point-based approach achieved high overall accuracy: 97.81 ± 0.55 % in Orco and 96.11 ± 0.64 % in Avisio. It demonstrated strong within-basin discrimination of peatland classes and generated detailed candidate maps, most of which are smaller than 1 ha. Performance declined under spatially separated cluster-based validation and varied among cluster combinations, indicating that transfer to spatially independent peatland systems is substantially more challenging. Models performed consistently better in Orco than in Avisio, likely reflecting differences in environmental conditions and peatland composition. Hydro-geomorphological variables were the most important predictors; Sentinel-2 variables contributed more strongly in Avisio and Sentinel-1 variables had limited importance. Overall, the framework accurately identified candidate alpine peatlands across geographically distinct study regions, while showing that performance strongly depends on the reference data spatial structure.