Capturing constraints on boreal gross primary productivity using the remote sensing-based CAN-TG model.

In response to the limited number and distribution of in-situ carbon flux observations, remote sensing-based methods are increasingly relied upon for the estimation of Gross Primary Productivity (GPP) at regional to global scales. These remote sensing-informed estimates are commonly derived through process-based modelling frameworks which prescribe functional relationships between model inputs and target GPP. Across highly heterogeneous landscapes like the Canadian boreal, these parameters are difficult to constrain and often site-specific. Recent work has determined that parameterization alone may not improve model performance, instead requiring additional model inputs to capture the complex drivers of vegetation productivity across land cover types. In response to these challenges, we applied the remote sensing-based CAN-TG framework to estimate boreal GPP, leveraged through a random forest (RF) machine learning approach that does not assume linear or functional relationships between input variables and productivity. Stratified by land cover, fire disturbance history, and topography, models were assessed for their ability to capture reference GPP from NASA's complex, process-based Soil Moisture Active Passive (SMAP) GPP product. Across all boreal strata, model r 2 values ranged from 0.93 to 0.96, demonstrating that the variability in substantially more complex models can be successfully captured using a simple, interpretable remote sensing-based framework. Through the addition of remote sensing variables capturing freeze/thaw and soil moisture dynamics to surface temperature and greenness, the CAN-TG model demonstrated an improved ability to capture GPP compared to a benchmark GPP model. Seasonal RF models across key boreal land cover, fire disturbance history and topographic strata further demonstrated varying and complex non-linear relationships between model variables and GPP. Spring and fall models generally outperformed winter and summer models, reaffirming model strengths whilst also highlighting remaining uncertainty and areas for future model improvement. • CAN-TG estimates of GPP explain 93–96 % of variation in reference GPP across strata. • Inclusion of soil moisture and freeze/thaw estimates improved model performance. • Drivers of boreal GPP have non-linear effects on model estimates. • Variable contribution to model training varies across boreal strata and seasons.