Improving species distribution models for stream networks by incorporating spatial autocorrelation in multi-sourced datasets: An assessment of Idaho giant salamander status and future risk

Fundamental to species conservation efforts is the development of accurate distribution models, but doing so is challenging for many stream organisms, where limited funding often necessitates the compilation of incidental observations from multiple sources, which lack an overall sampling design and may be spatially clustered. We demonstrate the application of specialized spatial-statistical-network models (SSNMs), which incorporate autocorrelation among observations and significantly outperform non-spatial models when used to develop distribution models for the Idaho giant salamander (IGS; Dicamptodon aterrimus). The study was located in the Rocky Mountains in west-central North America. We compiled a comprehensive presence-absence dataset for IGS from previous studies, natural resource agencies, museum collections, and new surveys and linked these data to geospatial habitat covariates. The dataset was modeled using a suite of candidate SSNMs and results were compared to generalized linear models (GLM). The top-ranked models were used to predict range-wide IGS occurrence probabilities for scenarios that represent historical baselines and futures associated with two model covariates (water temperature and riparian tree canopy density) that were changing with environmental trends in the study area.** The classification accuracy of salamander observations was higher with SSNMs than GLMs (90.8% versus 63.2%) and the spatial models identified fewer significant habitat relationships, which simplified model interpretation. Baseline range estimates from the models were similar (13,090–14,114 stream km) and both predicted small range expansions (2.0% to 24.8%) with warming because many streams were sub-optimally cold for IGS. However, these expansions could be partially offset in future scenarios which included decreases in riparian canopy density. **SSNMs significantly improve distribution models for stream organisms by incorporating spatial autocorrelation and provide an inexpensive means of developing new information from existing datasets. This incentivizes aggregation of datasets, which may be further leveraged to create efficient monitoring and inventory programs using the spatially-explicit outputs from SSNMs.

Methods
A presence-absence dataset for Idaho giant salamander that consisted of 707 unique sampling locations was collected using electrofishing and eDNA surveys. Many of the surveys were aggregated from existing sources such as previous peer-reviewed studies, grey literature reports, state and federal agency databases, and natural resource museum records. The survey locations were attached to reaches within stream networks across the species range, linked to geospatial habitat covariates, and processed using the open-source SSNbler R package into a landscape network object suitable for spatial-stream-network model analysis using the SSN2 R package.