Forest ecosystem restoration potential unravelled through machine learning algorithms

Decades of scientific development have led to a specialisation of environmental analyses, hampering the recognition of forest ecosystem restoration. However, an integrated approach is required to better understand the trajectory of forest recovery. In this study, we aimed to evaluate potential comprehensive approaches for forest restoration assessment by considering 184 environmental features, including geographical information, tree dimensions and biomass, soil abiotic parameters, microclimate, and soil microbial taxa and functions. All these characteristics constituted the dimensions of the ecosystem hypervolume and its emerging properties, related to naturally regenerating Acacia mangium plantations. Through unsupervised and supervised machine learning algorithms, we recognised the known time lag between the modification of ecosystem properties and the macroscopic aspects, such as the aboveground biomass. While the sectorial analysis highlighted the recovery of numerous characteristics by the 10 ^th year after plantation establishment (e.g., tree biomass), the analysis integrating numerous ecosystem parameters highlighted the sequential similarity along the chronosequence, with only the 24-year-old plantation approaching recovery as measured in an adjacent remnant forest. The presence of keystone parameters (environmental variables with disproportionate effect on the ecosystem restoration relative to their value or variance) and the lagged response of ecosystem properties to drivers of environmental changes call for a more comprehensive approach to assess the achievement of restoration goals. Together with recent efforts to include machine learning analyses to answer ecological questions, our study brings unprecedented evidence on how leveraging advances in this analytical field can prompt a better understanding and management of recovering forests, considering multiple intertwined environmental factors.