Soil erosion degrades soil and water resources in tropical protected areas. We mapped RUSLE-predicted soil loss in Ir. H. Djuanda Forest Park, West Java, Indonesia, and quantified the contribution of soil map units (SMUs) to spatial variation in model inputs. Rainfall erosivity was derived from 2016–2025 CHIRPS precipitation using the modified Fournier index and Arnoldus relation, giving a period-mean R of 13,825.61 MJ mm ha−1 h−1 yr−1. Soil erodibility was derived from a 1:50,000 soil map and SoilGrids organic carbon, the topographic factor from the DEMNAS grid using the Desmet–Govers formulation, and cover from SPOT-6 imagery (overall accuracy 82.5%, kappa 0.77). Cells with at least 300 contributing cells were treated as concentrated-flow cells and excluded from the hillslope calculation. The baseline RUSLE × SMU intersection contained 493.32 ha of valid estimates, with 93.36% below 385.14 t ha−1 yr−1 and a maximum predicted soil loss of 8,184.15 t ha−1 yr−1. SMU-4 contained 97.0% of the 32.74 ha above 385.14 t ha−1 yr−1. The area-weighted variance of ln(K) was 0.16% of that of ln(C), indicating substantially greater spatial contrast from land cover than from SMU-based K classes. Sensitivity testing showed that concentrated-flow treatment strongly affected the upper tail, whereas median LS changed by less than 10% between the unmasked and 75-cell scenarios. Without independent erosion measurements, the results are interpreted as modelled soil-loss potential and spatial screening information rather than measured erosion rates.