An unsupervised machine learning approach for enhancing forest structure estimation in uneven-aged stands

Accurate estimation of forest structural variables is crucial for improving predictions of ecosystem dynamics, assessing biodiversity, and promoting sustainable forest management, particularly in the context of climate change. This study evaluates the performance of traditional and unsupervised clustering methods in estimating key structural metrics - diameter at breast height (DBH), tree height (H), and basal area (BA). While traditional grouping methods, based on uniform H or DBH categories, have been widely used, they often fail to capture the inherent heterogeneity of uneven-aged forest stands. In this study, we explore the application of advanced unsupervised clustering methods - Principal Component Analysis (PCA), Gaussian Mixture Models (GMM), Hierarchical Clustering (HC), and Density-Based Clustering (DBSCAN) - to enhance the estimation of forest structural variables. Using 30 years of inventory data from uneven-aged Norway spruce-dominated stands in southern Finland, combined with simulations from the process-based model PREBAS, we evaluate the effectiveness of these clustering methods against traditional approaches. Results reveal that unsupervised methods consistently outperform traditional grouping techniques across all metrics. For BA, DBSCAN achieved an r 2 of 0.90, while PCA and GM reached r 2 values of 0.84 for H. Unsupervised methods also exhibited significantly lower RMSE and bias, particularly in capturing complex processes such as ingrowth, mortality, and thinning events. By contrast, traditional methods struggled with these dynamics, often overestimating key metrics and failing to represent stand-level heterogeneity. Species- and site-specific analyses highlighted variability in performance, with Pinus sylvestris exhibiting the lowest estimation accuracy due to sparse regeneration, whereas Norway spruce consistently showed robust recruitment. These findings underscore the value of unsupervised clustering in addressing the challenges posed by uneven-aged forests, providing more reliable inputs for process-based models and facilitating improved predictions of forest dynamics. This study highlights the potential of integrating advanced clustering techniques into forest modelling to refine structural estimations, support biodiversity conservation, and enhance carbon sequestration strategies. By capturing the complexity of forest stands more effectively than traditional methods, these approaches pave the way for innovative solutions to sustainable forest management under global environmental change.