A0382
Title: A cross-validation framework for log-Gaussian Cox processes with presence-only data in marine ecology
Authors: Daniele Poggio - Politecnico di Torino (Italy) [presenting]
Gian Mario Sangiovanni - Sapienza University (Italy)
Gianluca Mastrantonio - Politecnico of Turin (Italy)
Giovanna Jona Lasinio - Sapienza University of Rome (Italy)
Daniele Ventura - Sapienza University (Italy)
Edoardo Casoli - Sapienza University (Italy)
Stefano Moro - Stazione Zoologica Anton Dohrn (Italy)
Abstract: Building on a recent study, a novel implementation of a novel cross-validation strategy is proposed for model selection within the framework of log-Gaussian Cox processes (LGCP), designed to enhance predictive performance assessment. The approach implements a k-fold cross-validation scheme based on a spatial thinning mechanism to define the folds. Model comparison is performed by computing posterior distributions of raw residuals across a partitioned spatial domain and summarizing them using the continuous ranked probability score (CRPS), yielding a single metric per model. The framework is computationally efficient and suitable for integrating heterogeneous presence-only data, with particular relevance to marine ecology. Its application is demonstrated in modeling the spatial distribution of Holothurians recorded during nine survey campaigns (2022/2024) around Giglio Island, Italy. The surveys integrated high-resolution photogrammetry with diver-based visual censuses, resulting in variable detection probabilities across habitats, especially in Posidonia Oceanica meadows. To address this complexity, LGCP models are adopted with a shared spatial Gaussian process component, capturing habitat structure, environmental covariates, and temporal variability. Multiple model specifications are evaluated, each incorporating different sets of spatial covariates reflecting habitat and geomorphological features, and the most predictive model is identified.
