A budget-constrained input state-space compression framework for scaling process-based forest growth models: A 3-PG case study in Ontario

: Large-scale applications of process-based forest growth models are often constrained by the large number of forest conditions that must be represented and the computational burden of simulating them. To address this challenge, we developed a framework that replaces the original forest population with a smaller set of representative stands to simulate biomass across forest areas spanning an entire province, using the 3-PG forest growth model in Ontario, Canada, as a case study. Two million stands of forest resource inventory data from 41 forest management units in Ontario were used to implement the proposed framework. Species-specific 3-PG parameters were estimated for 14 dominant tree species, together accounting for more than 97% of the managed forest area, using constrained maximum a posteriori (MAP) estimation. Mixed stands were represented by their three most abundant species and their normalized composition proportions. The stand population was then stratified by dominant species, and climate and stand attributes were jointly clustered to identify representative stands. Two compression strategies were evaluated: K -means clustering based on climate and stand variables, and an XGBoost-guided version that additionally incorporated predictions of 3-PG biomass. Across the tested subset sizes, retaining more representative stands generally improved accuracy but also increased computation time, demonstrating the tradeoff between simulation accuracy and computational efficiency. At comparable compression levels, XGBoost-guided clustering generally achieved lower root mean square error (RMSE) and relative root mean square error (RRMSE) and higher coefficient of determination ( R 2 ) than pure K -means clustering. This advantage was particularly evident for mixed stands: pure K -means primarily distinguished stands by dominant species and continuous climate and stand attributes, whereas XGBoost learned biomass patterns associated with multiple species and their composition proportions. Beyond approximately 2,000 representative stands, retaining additional stands produced comparatively small reductions in prediction error, while computation time continued to increase. Therefore, approximately 2,000 representative stands provided a practical balance between simulation accuracy and runtime. Using approximately 2,000 representative stands, the framework reproduced the uncompressed 3-PG estimates of total tree biomass across Ontario’s managed forests, calculated as the sum of stem, foliage, and root biomass, in about 790 s , compared with an estimated 108,000 s required to simulate every stand. These results showed that the proposed framework can substantially reduce computational cost while preserving a high level of agreement with the full simulations, thereby providing a practical basis for large-scale biomass assessment and forest management analysis.