Improving MaxEnt reliability with multi-criteria analysis and site weighting: A case study on Caulerpa cylindracea

The rapid spread of invasive alien species (IAS) underlines the urgent need for predictive modelling to accurately forecast future spread under climate change. Correlative ecological niche models (ENMs) serve this purpose, but often struggle with sampling bias, overfitting and uncertainty quantification and rely on a certain degree of subjectivity, which limits their reliability. We have developed an improved ENM framework using MaxEnt that integrates standard and site-weighted performance metrics into a multi-criteria decision process to select models that provide a better balance between explanatory power and transferability. To detect overfitting, we introduced delta metrics that measure the dependence of model performance on the weighting scheme. Furthermore, we performed a sensitivity analysis to quantify the classification uncertainty associated with different probability thresholds. We test this approach on Caulerpa cylindracea, one of the most dangerous IAS in the Mediterranean. The selection approach identified an optimal model that showed high and stable performance across all weighting schemes and multiple validation datasets. Both the annual and scenario-based projections show a general shift towards lower habitat suitability, with a statistically significant negative trend in high suitability areas. The decline suggests that the future spread of C. cylindracea is likely to be limited to currently invaded areas, assuming no adaptation. The high reliability of the model is supported by the extremely low extrapolation risk (0.21% of the predictions under ‘strict extrapolation’) and the agreement of the response curves with the known ecophysiology of the species. However, the sensitivity analysis of threshold selection shows a non-uniform pattern of classification uncertainty that appears to be related to invasion stage, with the greatest variability observed in recently invaded regions such as the Northern Adriatic. Practical Implication: This framework provides robust predictions of invasion risk while explicitly quantifying uncertainty. The selection procedure, through delta metrics, explicitly includes model generalisation ability in the ranking process. The annual suitability projections enable temporal prioritisation of control efforts. Threshold sensitivity analysis identifies areas that require more conservative management approaches, improving the quality and cost-effectiveness of intervention strategies.