Species Distribution Modeling with Large Occurrence Databases: From Observer Bias Correction to Data Integration Strategies

Ing. Petr Balej

Disertační práce

Abstract:

The rapid expansion of large-scale species occurrence databases, driven by citizen science, presents unprecedented opportunities for biodiversity research. However, reliability is limited by pervasive sampling bias, positional uncertainty, and heterogeneous observation processes. This dissertation develops and tests methods to improve the reliability of species distribution models (SDMs), evaluated across 108 Central European bird species. Core contributions are threefold. First, it synthesizes how sample size, positional uncertainty, and sampling bias affect SDMs and how these effects depend on species ecology. Second, it develops the Presence-Weighted Observer-Oriented Approach (PW-OOA), a method that corrects sampling bias by explicitly modeling individual kernel density surfaces derived from each observer’s target group occurrence records. These observer-specific surfaces are then summed into a species-specific bias raster, with each observer’s contribution weighted according to their relative presence in occupied pixels for the focal species. Background points are subsequently sampled from this species-specific surface, ensuring that the simulated background distribution reflects species-relevant observer behavior rather than global observer patterns. The method is applied using data from the Czech Species Occurrence Database (CSOD), which provides comprehensive national-scale occurrence records. We evaluate PW-OOA against spatial thinning and target-group approaches using an independent presence–absence survey (PAD), which demonstrates the evaluation paradox, where internal cross-validation can favor uncorrected models while independent validation favors bias-correction approaches. Third, the battle of biases reveals the consequences of cross-border data integration by comparing models built from CSOD to models integrating global data from the Global Biodiversity Information Facility (GBIF), including eBird records for which the standard positional uncertainty attribute is not provided in GBIF. It quantifies the expansion of sampled environmental conditions using the Sampling Breadth Index (SBI) as an indirect indicator of potential niche truncation and shows that broader environmental coverage does not necessarily translate into improved country-level predictive performance. Overall, PW‑OOA consistently outperforms other tested bias correction approaches. Moreover, this thesis proves that independent validation and explicit treatment of sampling bias are necessary for reliable SDMs based on unstructured opportunistic occurrence data, and that the trade-off between potential niche truncation and heterogeneous sampling bias should be evaluated when integrating datasets.

Keywords

Observer-Oriented Approach, Data Integration, Species Distribution Modeling, Sampling Bias, Citizen Science, Niche Truncation, Background Points, Spatial Point Process

Jazyk práce: angličtina

Datum vytvoření / odevzdání či podání práce: 11. 2. 2026

Obhajoba závěrečné práce

• Vedoucí: prof. Ing. Petra Šímová, Ph.D.
• Oponent: Tomáš Václavík, externi, Angela Lausch, externi

Česká zemědělská univerzita v Praze

Fakulta životního prostředí

Doktorský studijní program:
Aplikovaná a krajinná ekologie