Predicting cultivation probability to target monitoring in smallholder landscapes: Proof of concept from dispersed medicinal plant cultivation in Taihe County, China

Dispersed medicinal plant cultivation by smallholder farmers presents a monitoring challenge that differs fundamentally from conventional agricultural monitoring: fragmented fields, mixed spectra, and canopy occlusion make direct detection difficult, while exhaustive field surveys are prohibitively costly. Rather than detecting every plot, we combined Random Forest modeling with geographical detector analysis to predict cultivation probability from environmental covariates, transforming monitoring from an open-ended search into targeted, spatially guided effort. Using Plantago asiatica L. in Taihe County, China, a 300 years of cultivatd species as a model system. Random Forest achieved stable predictive performance under 5-fold cross-validation (AUC = 0.977 ± 0.022; coefficient of variation = 2.3%). Geographical detector analysis revealed a clear interaction hierarchy. Elevation (q = 0.620) and land use (q = 0.464) as the dominant drivers, and elevation functioned as a universal amplifier, producing consistently strong synergistic effects with any co-occurring factor, in contrast to the highly variable interactions among soil physical properties (CV = 2.8% versus 21.3%). Risk detection further delineated optimal environmental ranges for each factor. Cross-validation between geographical detector and Random Forest importance rankings (Spearman's ρ = 0.664) confirmed the robustness of these mechanistic insights. To inform management, we derived a three tier monitoring priority zoning from the probability surface, in which the high-priority zone, occupying only 9.67% of the county yet comprising over 91% cropland and artificial surfaces, serves as the primary monitoring stratum. This design theoretically reduces annual monitoring workload by 80–90% compared to exhaustive surveys. The approach relies on open-access data and open-source software, providing a template potentially adaptable to other crops and regions with long term cultivation histories, though operational validation remains necessary. By resolving the factor interaction hierarchy, this study advances both mechanistic understanding of cultivation patterns and the design of targeted monitoring strategies in smallholder-dominated landscapes.