Accounting for spatial varying sampling effort due to accessibility in Citizen Science data: A case study of moose in Norway

Abstract

Citizen Scientists together with an increasing access to technology provide large datasets that can be used to study e.g. ecology and biodiversity. Unknown and varying sampling effort is a major issue when making inference based on citizen science data. In this paper we propose a modeling approach for accounting for variation in sampling effort due to accessibility. The paper is based on a illustrative case study using citizen science data of moose occurrence in Hedmark, Norway. The aim is to make inference about the importance of two geographical properties known to influence moose occurrence; terrain ruggedness index and solar radiation. Explanatory analysis show that moose occurrences are overrepresented close to roads, and we use distance to roads as a proxy for accessibility. We propose a model based on a Bayesian Log-Gaussian Cox Process specification for occurrence. The model accounts for accessibility through a distance sampling approach. This approach can be seen as a thinning process where probability of thinning, i.e. not observing, increases with increasing distances. For the moose case study distance to roads are used. Computationally efficient full Bayesian inference is performed using the Integrated Nested Laplace Approximation and the Stochastic Partial Differential Equation approach for spatial modeling. The proposed model as well as the consequences of not accounting for varying sampling effort due to accessibility are studied through a simulation study based on the case study. Considerable biases are found in estimates for the effect of radiation on moose occurrence when accessibility is not considered in the model.

Consequences of ignoring variable and spatially-autocorrelated detection probability in spatial capture-recapture

“…Situations where entire regions of the study area remain unsampled are fairly common in monitoring studies, where the data is gathered opportunistically, either in part or as a whole (Conn et al 2017; Altwegg and Nichols 2019; Sicacha-Parada et al 2020). Unstructured sampling can augment information, and thus improve population inferences about rare or elusive species (Thompson et al 2012; Tenan et al 2017; Sun et al 2019; Bischof et al 2020a).…”

“…Opportunistically-collected data obtained as part of public surveys (e.g. citizen-science data) are sometimes integrated into monitoring programs as this allows investigators to sample areas at unprecedented scales with lower costs and the added benefit of public involvement in management and conservation practices (Altwegg and Nichols 2019; Bischof et al 2020a; Sicacha-Parada et al 2020). To minimize and account for variation in detection probability in such sampling schemes, volunteers should be encouraged to visit all habitats within the study area, reduce variability in observer proficiency by providing standardized training, and collect data on potentially relevant covariates (Altwegg and Nichols 2019).…”

“…Survey effort may also vary across the study area unbeknownst to the investigator. For example, many large-scale monitoring programs combine structured sampling with unstructured data collection methods to increase the spatial extent and sampling intensity, and involve members of the public in the process (Thompson et al 2012; Conn et al 2017; Altwegg and Nichols 2019; Bischof et al 2020a; Sicacha-Parada et al 2020). Data from unstructured sources introduce unknown spatial heterogeneity in detection probability in SCR studies.…”

Species-Habitat Associations: Spatial data, predictive models, and ecological insights

“…These issues are particularly relevant to data collected haphazardly or opportunistically as part of museum collections or citizen science programs (Fithian et al, 2015). In such cases, it will often be fruitful to model the data as a thinned point-process model (Warton, Renner, & Ramp, 2013;Dorazio, 2014;Fithian et al, 2015;Sicacha-Parada et al, 2020). Specifically, we may think of the observed data as being generated by 2 processes: 1) the biological process model (as discussed in Chapter 2) written in terms of an Inhomogeneous Poisson Point-Process, and 2) an independent observation model that "thins" the data, specified using a separate function p(s) that describes the likelihood of observing an individual, conditional on that individual being present at location s. Typically, p(s) will also be written as a log-linear function of spatial (and potentially non-spatial) covariates (z 1 , z 2 , .…”

Ecological impact assessment of climate change and habitat loss on wetland vertebrate assemblages of the Great Barrier Reef catchment and the influence of survey bias

“…Consequently, the data were skewed toward cities, rather than representative of wetland distribution, which means that data for wetlands that are more distant from urban centers were not available, which severely limits the ability to run species distribution modeling. It is not uncommon for citizen science programs, outside a structured or randomized assessment program, to yield greater species detection rates closer to home (Sicacha-Parada et al, 2020;Tulloch & Szabo, 2012). Not assessed here is also the potential impact of land accessibility and the potential influence of small towns between cities, which can both be potential sources of noise.…”