How to Optimize Multispecies Set Predictions in Presence-Absence Modeling ?

Species distribution models (SDMs) commonly produce probabilistic occurrence predictions that must be converted into binary presence-absence maps for ecological inference and conservation planning. However, this binarization step is typically heuristic and can substantially distort estimates of species prevalence and community composition. We present MaxExp, a decision-driven binarization framework that selects the most probable species assemblage by directly maximizing a chosen evaluation metric. MaxExp requires no calibration data and is flexible across several scores. We also introduce the Set Size Expectation (SSE) method, a computationally efficient alternative that predicts assemblages based on expected species richness. Using three case studies spanning diverse taxa, species counts, and performance metrics, we show that MaxExp consistently matches or surpasses widely used thresholding and calibration methods, especially under strong class imbalance and high rarity. SSE offers a simpler yet competitive option. Together, these methods provide robust, reproducible tools for multispecies SDM binarization.

💡 Research Summary

This paper addresses a critical yet under‑explored step in species distribution modelling (SDM): the conversion of continuous probability outputs into binary presence‑absence maps. While many ecological applications require binary decisions (e.g., range mapping, community composition, protected‑area planning), the standard practice of applying arbitrary thresholds or calibration procedures can introduce substantial bias, especially for rare species and highly imbalanced datasets.

The authors introduce two novel, unsupervised binarization frameworks that directly maximize a chosen evaluation metric without requiring any external calibration data. The first method, MaxExp (Maximum Expected Score), formulates the problem as maximizing the expected value of a similarity function U (e.g., F1‑score, Jaccard, True Skill Statistic) between the predicted assemblage and the true (unknown) assemblage at a given site. Under two reasonable assumptions—(A1) the metric depends only on TP, FP, FN, TN, and (A2) species presences are independent—the expected score can be expressed solely in terms of the marginal probabilities of each species. Consequently, the optimal prediction reduces to selecting the top‑k species with the highest predicted probabilities, where k is the number of species to be predicted at that site. The authors derive a closed‑form objective for k and show that it can be solved in O(N³) time (or O(N²) for specific scores such as F1 and Jaccard), making the approach tractable even for hundreds of species.

The second method, Set Size Expectation (SSE), simplifies MaxExp by estimating k from the expected species richness (the sum of predicted probabilities) and then selecting the k most probable species. This reduces computational complexity to O(N log N) while retaining competitive performance.

To evaluate the methods, the authors conduct three case studies spanning marine reef fish (≈150 species), tropical reef fish (≈300 species), and avian/insect citizen‑science data (≈500 species). They compare MaxExp and SSE against a suite of reference binarization techniques, including traditional thresholding (Youden Index, fixed 10 % threshold), calibration‑based methods (Platt scaling, isotonic regression), and a recent conformal prediction approach. Performance is assessed using sample‑averaged metrics (F1, Jaccard, TSS) and richness error. Across all scenarios, MaxExp consistently achieves the highest scores, particularly excelling in preserving rare‑species detections while maintaining overall accuracy. Calibration‑based methods suffer from over‑fitting when validation data are scarce, and simple thresholds tend to inflate false negatives for low‑prevalence species. SSE, while slightly less optimal than MaxExp, delivers comparable results with far lower computational demand, making it suitable for large‑scale or real‑time applications.

A key advantage of both frameworks is that they require only the probabilistic outputs already generated by the SDM; no additional presence‑absence observations are needed for binarization, thereby avoiding the risk of over‑fitting and reducing data collection burden. The authors provide fully reproducible code (GitHub) and open data (SeaFile), encouraging immediate adoption.

In conclusion, MaxExp offers a theoretically sound, metric‑driven solution for multispecies binarization that outperforms existing heuristic and calibration approaches, while SSE provides a lightweight alternative with competitive accuracy. The work paves the way for more reliable community‑level predictions in conservation planning, climate‑impact assessments, and biodiversity monitoring. Future extensions may relax the independence assumption, incorporate species interactions, or jointly optimize multiple objectives such as richness and habitat suitability.