Integrating deep learning and spatial statistics in marine ecosystem monitoring
Environmental and Ecological Statistics · 2026-06 · 33 (2) · 1161-1183
DOI: 10.1007/s10651-026-00732-7 ↗Kurz gefasst
Fachartikel in der Zeitschrift Environmental and Ecological Statistics, online erschienen am 15.05.2026. Verfasst von Gian Mario Sangiovanni und 4 weiteren Autor:innen. Die Arbeit nennt 66 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Sapienza University of Rome, Università degli Studi di Roma La Sapienza.
Abstract (Originalsprache)
In ecology, photogrammetry is a crucial method for efficiently collecting non-destructive samples of natural environments. When estimating the spatial distribution of animals, detecting objects in large-scale imagery becomes essential. Deep learning-based object detection enables large-scale analysis but introduces uncertainty because detection probability depends on environmental and observational factors. To address detection bias, we model the distribution of benthic sea cucumbers in an area of the Italian Tyrrhenian coast near Giglio Island using a Thinned Log-Gaussian Cox Process (LGCP). The main objective of this study is to develop a modular framework that integrates a deep learning-based object detector with a spatial point process to correct for systematic undercounting in marine surveys. We employ the YOLOv11 architecture to automate the identification of individuals. We assume that a true underlying intensity describes the actual population, while the observed detections correspond to a subset of individuals that are not always detected, leading to a degraded intensity. The detection process is modeled through a probabilistic function that depends on environmental and observational factors, including local density, network confidence scores, and object size. Manual annotations are used as a benchmark, and we compare the thinned LGCP with an unthinned model fitted to the deep learning-based object detections. The proposed approach reduces bias in spatial intensity estimation and improves agreement with manual surveys (average Pearson residual 6.150 vs. 6.694 and average raw residual 1.979 vs. 2.092). By explicitly accounting for detection uncertainty, the framework increases the reliability of large-scale benthic monitoring, supporting habitat assessment and evidence-based marine conservation.
Förderung
- Sapienza University of Rome
- Università degli Studi di Roma La Sapienza
Zitieren
APA
Sangiovanni, G. M., Mastrantonio, G., Pollice, A., Ventura, D., & Lasinio, G. J. (2026). Integrating deep learning and spatial statistics in marine ecosystem monitoring. Environmental and Ecological Statistics, 33(2), 1161-1183. https://doi.org/10.1007/s10651-026-00732-7
Vancouver
Sangiovanni GM, Mastrantonio G, Pollice A, Ventura D, Lasinio GJ. Integrating deep learning and spatial statistics in marine ecosystem monitoring. Environmental and Ecological Statistics. 2026;33(2):1161-1183. doi:10.1007/s10651-026-00732-7
BibTeX
@article{Sangiovanni2026,
author = {Sangiovanni, Gian Mario and Mastrantonio, Gianluca and Pollice, Alessio and Ventura, Daniele and Lasinio, Giovanna Jona},
title = {Integrating deep learning and spatial statistics in marine ecosystem monitoring},
journal = {Environmental and Ecological Statistics},
year = {2026},
volume = {33},
number = {2},
pages = {1161--1183},
doi = {10.1007/s10651-026-00732-7}
}
EUVIMED