Toward a deeper understanding of recovery trajectories in forest ecosystem restoration: a machine learning approach

Abstract Introduction The lack of integrated monitoring approaches and indicators hampers the understanding of forest restoration trajectories, as the recovery process involves multiple aspects of the ecosystem, which can respond differently to environmental changes due to their interactions and synergies. Objective We aimed to demonstrate that a highly integrated approach and adoption of machine learning methods can provide a deeper understanding of forest recovery trajectory, in a context where the subcomponents may vary at different rates. Specifically, we considered 184 ecosystem attributes, including geographical information, tree community characteristics, soil fertility, and soil microbial taxa and functions, to study the trajectory of forest restoration in former Acacia mangium plantations. These ecosystem characteristics encompass multiple dimensions of the ecosystem hypervolume, from which ecosystem properties emerge. Methods We analyzed and processed previously collected data on the spatial location, tree community, aboveground biomass, and soil microbial assemblages through unsupervised and supervised machine learning algorithms. Our approach is based on space‐for‐time substitution rather than surveys over time within sites. Results The analysis integrating the 184 ecosystem factors revealed that landcover types were similar only to the sequence nearest in age (e.g., 10‐year‐old plantation similar to the 2‐ and 24‐year‐old plantings), even when individual parameters recovered at different rates.