Bayesian modelling of species-environment relationships for partially observed data. - TEL - Thèses en ligne Accéder directement au contenu
Thèse Année : 2022

Bayesian modelling of species-environment relationships for partially observed data.

Modélisation Bayésienne des relations espèces-environnement lorsque les données sont partiellement observées.

Résumé

Characterising how species respond to its environment is of central interest in ecology. Species-environment relationships (SERs) are studied in many topics, for instance, in community ecology, in species distribution modelling, and to guide conservation or management actions. Statistical models fitting species distribution data (e.g., presence/absence or counts) to environmental data (e.g., temperature) are often used to estimate SERs. Standard models assume that data is representative of the SER. However, data can represent only a partial description of the SER. In this work, we investigated the effects on modelling SERs of three kinds of partially observed data:1) Partially observed response data, e.g., species sampled occurrences may only represent a partial observation of the occupancy status due to missing species present (i.e., imperfect detection).2) Partially observed environmental data, e.g., environmental descriptors may represent averaged conditions at a coarser spatial scale than the one at which SER is studied (i.e., area-to-point spatial misalignment).3) Partially observed relationship, e.g., the gradient of environmental conditions that describe the SER are not entirely surveyed (i.e., truncated gradient).Hierarchical Bayesian Models, allowing multi-species inferences and disentangling ecological from observational processes, have been developed and tested in three case studies, each involving a particular type of partially observed data.In the first case study, we emphasized that even a robust sampling design that involves multiple sampling replicates and detection techniques can lead to species detection probabilities lower than one in an insect community. We then advocated for the use of Multi-Species Occupancy Models to account for imperfect detection in insect studies. In the second case study, we showed how using area-to-point misaligned covariate can flatten SERs estimated by generalized linear models and how fitting a Berkson error model can lower the bias. In the third case study, we developed a hierarchical model that explicitly estimates optimum shifts. By constraining estimated SERs to concave shapes (following ecological theory), the new model improved estimates relative to past methods, especially in the case of truncated gradients.
Comprendre et prédire la distribution des espèces dans leur environnement est enjeu majeur en écologie. De nombreux outils statistiques ont été développés et sont appliqués pour étudier la distribution des espèces et leurs changements. Dans cette thèse nous nous intéressons à l'utilisation des modèles hiérarchiques bayésien dans ce contexte. Plus particulièrement nous cherchons à étudier l'apport de l'intégration de connaissances écologiques dans ces modèles. Trois applications écologiques serviront d'appui : (1) Étude de la distribution d'une communauté d'Orthoptères soumis à une détection imparfaite, (2) Estimation du déplacement de l'optimum entre deux distributions unimodales, (3) Effet de la résolution d'échantillonnage sur la prédiction de la répartition de la palourde japonaise.
Fichier principal
Vignette du fichier
thesismourguiart.pdf (9.37 Mo) Télécharger le fichier
Origine : Version validée par le jury (STAR)

Dates et versions

tel-04085564 , version 1 (29-04-2023)

Identifiants

  • HAL Id : tel-04085564 , version 1

Citer

Bastien Mourguiart. Bayesian modelling of species-environment relationships for partially observed data.. General Mathematics [math.GM]. Université de Pau et des Pays de l'Adour, 2022. English. ⟨NNT : 2022PAUU3023⟩. ⟨tel-04085564⟩
104 Consultations
32 Téléchargements

Partager

Gmail Facebook X LinkedIn More