Forest canopies significantly decouple sub-canopy microclimates from regional free-air temperatures, which creates complex thermal regimes that drive ecohydrological processes such as evapotranspiration and snowmelt. However, standard weather data often fails to capture these site-specific variations, especially in complex terrain. This study introduces a modelling framework applied across British Columbia and Alberta to downscale macroclimate data into biologically relevant microclimate estimates. We synthesized an extensive dataset of near-surface sub-canopy temperatures (2005–2024) from a broad network of data loggers spanning British Columbia and Alberta. These in-situ observations were paired with elevation-adjusted monthly air temperature grids from ClimateNA to calculate site-specific seasonal temperature offsets. To predict these offsets, we utilized a machine learning-based multi-model ensemble approach (LightGBM, XGBoost, Boosted Regression Trees, and Random Forest) using a diverse suite of covariates, including fine-scale topographic indices, and canopy structure, and macroclimatic energy and water balance metrics, to capture the non-linear interactions governing microclimate heterogeneity. Beyond prediction, we employed SHAP (SHapley Additive exPlanations) values to interpret "black box" model behavior, spatially resolving the specific magnitude and direction of influence for diverse physical drivers across different seasons. Results indicate distinct seasonal shifts in these controls; while macroclimatic fluxes were the primary drivers of summer and fall temperature offsets, local topographic features exerted their strongest influence during winter and spring. Additionally, uncertainty decomposition revealed that algorithmic variability frequently outweighed parameter tuning uncertainty, highlighting the necessity of ensemble approaches over single-model reliance. These high-resolution offsets extend beyond hydrology to inform species distribution, agricultural suitability, and climate micro-refugia identification.