Finding Inter-species Associations on Large Citizen Science Datasets

TL;DR

This paper presents a scalable method for identifying species associations from large citizen science datasets by analyzing spatial overlaps and comparing them to a null model, effectively filtering out spurious correlations.

Contribution

The paper introduces a novel, efficient approach to detect likely inter-species associations in large-scale citizen science data, accounting for observer bias and spatial heterogeneity.

Findings

Successfully identified known insect host-plant relationships.

Detected associations between Yerba Santa Beetles and California Yerba Santa.

Efficiently analyzed approximately 10^8 species pairs on modest hardware.

Abstract

Determining associations among different species from citizen science databases is challenging due to observer behavior and intrinsic density variations that give rise to correlations that do not imply species associations. This paper introduces a method that can efficiently analyze large datasets to extract likely species associations. It tiles space into small blocks chosen to be of the accuracy of the data coordinates, and reduces observations to presence/absence per tile, in order to compute pairwise overlaps. It compares these overlaps with a spatial Poisson process that serves as a null model. For each species $i$, an expected overlap $\mu_i$ is estimated by averaging normalized overlaps over other species in the same vicinity. This gives a $z$-score for significance of a species-species association and a correlation index for the strength of this association. This was tested on $874,263$ iNaturalist observations spanning $15,975$ non-avian taxa in the Santa Cruz, California region ($\approx 4.68\times10^{6}$ tiles). The method recovers well-known insect host-plant obligate relationships, particularly many host-gall relationships, as well as the relationship between Yerba Santa Beetles and California Yerba Santa. This approach efficiently finds associations on $\sim10^{8}$ species pairs on modest hardware, filtering correlations arising from heterogeneous spatial prevalence and user artifacts. It produces a ranked shortlist of ecological interactions that can be further pursued. Extensions to this method are possible, such as investigating the effects of time and elevation. It could also be useful in the determination of microhabitats and biomes.