Ecology-informed symbolic machine learning: a methodological framework for classification of forest succession

Accurately classifying forest successional stages remains a major challenge in applied ecology due to the continuum of succession, ecological heterogeneity, and limited interpretability of many machine learning (ML) approaches. Prevailing models typically rely on black-box algorithms that, while accurate, often lack ecological transparency, limiting their practical use in restoration and regulatory contexts. Here, we introduce and evaluate an ecology-informed symbolic machine learning (EISy-ML) framework that integrates symbolic regression with adaptive ecological constraints, specifically monotonic biomass trajectories and structural complexity proxies, derived from allometric functions. Using field data from 467 plots in the Subtropical Atlantic Forest, Brazil, EISy-ML generated interpretable and biologically plausible equations for successional classification. Performance was benchmarked against eight standard ML classifiers using balanced accuracy, macro F1, Cohen's kappa, and Matthews correlation coefficient. EISy-ML achieved the highest test accuracy (0.899), F1 (0.905), Kappa (0.829), and MCC (0.803), with no statistically significant difference compared to the next best models. The symbolic framework offers substantial improvements in transparency, reproducibility, and ecological coherence over conventional approaches, enabling direct application in restoration monitoring and environmental auditing. These results validate the hypothesis that symbolic ML integrated with ecological constraints produces models that are both robust and operationally interpretable. Future research should extend EISy-ML validation to other biomes, incorporate temporal and functional trait data, and explore uncertainty-aware or fuzzy logic extensions for handling transitional successional states.