Forest ecosystem services (ES) emerge across multiple spatiotemporal scales, are context-dependent, and often display non-linear dynamics. Yet, ES assessments often rely on global statistical methods applied to predefined spatial units, limiting their ability to capture context-specific interactions at the scales at which they emerge. Addressing this gap is critical for integrating ES into planning and management, particularly in multifunctional forest landscapes. Here, we demonstrate how geographically weighted machine learning with adaptive kernels can be used to identify the spatial scales at which individual ES are most strongly expressed, reveal their underlying non-linear drivers, and capture interactions among ES operating at different scales. Focusing on montane and subalpine forests in Catalonia (northeast Spain), we assessed four forest ES indicators: biomass production, fire risk, mushroom production, and scenic beauty, empirically quantified based on the 399 permanent plots with the presence of Scots pine (Pinus sylvestris), black pine (Pinus nigra), mountain pine (Pinus uncinata), and European silver fir (Abies alba), re-measured in the three cycles of the Spanish National Forest Inventory spanning 1989 to 2015. We first identified dominant global drivers of each ES using the Random Forest (RF) algorithm and then applied Geographically weighted Random Forest (GRF) to capture spatial heterogeneity in ES-driver relationships. Local scales were defined as the spatial extents that maximized model performance. These scales were subsequently employed in spatially weighted non-parametric correlation analyses to assess ES interactions. Our results reveal scale and driver dependency: ES dominated by a single structural driver, such as biomass production and fire risk, operated at broader spatial scales and exhibited relatively consistent interaction patterns, whereas services shaped by multiple co-dominant drivers, such as mushroom production and scenic beauty, displayed localized, heterogeneous patterns. By employing geographically weighted adaptive kernels, our framework captured both global and local ES relationships, demonstrating its value for multi-scale assessments, that can effectively aid geographically oriented forest management planning. Future work will expand the observational domain, thereby enhancing machine learning applications for spatially explicit forest ES analyses.