Adaptive local maxima windows for tree detection: A point process perspective

The growing accessibility of Light Detection and Ranging (LiDAR) data brings out novel perspectives that are crucial for tracking forest growth and enhancing resource management amid climate change. Utilizing these data to propose decision-support tools involves a vital step of segmenting individual trees. A widely adopted class of methods for this step is known as Local Maxima algorithms, which, although unsupervised, rely on per-site and/or per-species hyperparameter tuning for optimal performance. In this work, we introduce a novel methodological framework grounded in point process theory to jointly model the data generation process and provide formal implementation guidelines for refining window size selection within the class of Local Maxima algorithms. This methodology can also be applied to incomplete plot measurements, alleviating a constraint noted in most data acquisition procedures. To ensure the reproducibility of the results and validate the practical application, we apply the proposed methodology in two cases: (i) a simulated dataset (made publicly available) and (ii) an open real dataset. The simulation study evaluates performance under spatial configurations that do not necessarily follow the assumed point process model used for window calibration, thereby assessing robustness to model misspecification. The method outperforms the baseline approaches in the simulation study for the detection task, and achieves F 1 -scores between 55% and 90% on real data. On average, it improves upon the second-best method by about 4%, with performance depending on a tree’s position within the canopy relative to its neighbors.