Predicting species abundance using machine learning approach: a comparative assessment of random forest spatial variants and performance metrics

For informed decision-making in biodiversity conservation and ecological management, accurate predictions of species abundance are essential. This study aimed to assess the predictive performance of random forest (RF) spatial variants in modelling species abundance distribution compared to standard RF, Poisson, their hybrid methods with ordinary kriging (OK), and the random generalised linear model (RGLM). Model performance in abundance modelling has rarely been quantified using a comprehensive index, except the existing single statistical indices. Therefore, modified Taylor diagrams were used to evaluate the model's overall ability to predict species abundance spatial patterns, taking into account abundance class and detection probability. An exponential correlation function was used to generate spatially correlated random effects with and without a quadratic term and two variation strengths. Species abundance class and the relationship between abundance and independent variables determine which RF spatial variant performs the best. Spatial RF variants outperform conventional modelling in terms of prediction accuracy and power, particularly when spatial autocorrelation and species detection probabilities are high. RF spatial variants were less precise for common species than RGLM and GLM-OK, which better predicted species abundance for low or no spatial autocorrelation cases. However, none of the models outperformed the others for all prediction goals, highlighting the need for combining performance metrics to evaluate species abundance distribution models. The study highlights the importance of model specification in ecological research and cautions against the use of RF algorithms as a black box.