Title: Improving endangered species assessments using the Automated Probabilistic Co‐Occurrence Assessment Tool
Abstract: Under the US Endangered Species Act (ESA), calculating the degree of spatial overlap among endangered species' geographic ranges and registered pesticide use patterns is required and critical to assessing and mitigating potential risks posed to the species in question. However, the dynamic nature of cropping practices, limited data availability for rare species geographic ranges, and the large number of active ingredients and endangered species that must be evaluated can make conducting these co-occurrence analyses challenging and time-consuming. In response to the difficulty of characterizing uncertainty and variability in risk assessments of all types, the National Research Council recommends the use of probabilistic analysis methods (National Research Council, 2013). In the case of endangered species assessments for pesticide registration, researchers have applied probabilistic methods to produce crop footprints (spatial representations of agricultural pesticide use sites), which account for crop rotations and mapping uncertainty (Budreski et al., 2016), and to species distribution models (SDMs) (Phillips et al., 2017), which map variability in habitat quality across species geographic ranges based on species location records and associated environmental data. Research shows that using these probabilistic methods in co-occurrence analyses can yield significantly different results relative to commonly used deterministic methods (Richardson et al., 2022), especially for cases in which probabilistic SDMs identify suitable habitat outside of the previously delineated ranges. Yet, the problem of the large quantity of co-occurrence analyses potentially required for pesticide registration remains and is compounded by the greater effort required to use these probabilistic methods relative to deterministic methods. probabilistic pesticide usage raster data sets (gridded images composed of pixels, in this case measuring 30 m across) that comprise over 90% of cultivated acreage in the continental United States (i.e., soybeans, corn, wheat, cotton, rice, alfalfa); watershed-scale probabilities of pesticide occurrence in aquatic environments; probabilistic SDMs for both terrestrial and aquatic species; and reports on the spatial co-occurrence between pesticide use and species ranges in terrestrial and aquatic environments. Probabilistic pesticide use footprints generated with APCOAT account for variability in cropping practices, uncertainty in crop mapping, and spatial and temporal variability in pesticide usage. Modeling these factors begins with the national probabilistic crop footprint rasters included with APCOAT. In these rasters, cropping practice variability is represented by calculating the probability of crop presence over six years using the USDA Cropland Data Layer (CDL). Uncertainty in crop mapping is also incorporated using the principles of Bayes' Theorem, which is used to update probabilities by incorporating additional evidence. In this case, the additional evidence includes the annual state-level accuracy of the CDL by crop, its agreement with the National Land Cover Database (NLCD), and agreement with crop acreage reported to the National Agricultural Statistics Service (NASS). Spatial and temporal variability in pesticide usage on these crops is incorporated by using time series of crop acreage derived from the CDL, usage estimates for 452 commonly used pesticides (Wieben, 2019), and pesticide application rates provided by the user to generate state-level statistics of the percent crop treated (PCT). The user may also generate the PCT statistics at the resolution of county or Crop Reporting District if they are able to provide the relevant pesticide usage estimates. Once the PCT statistics are generated, they are applied to the probabilistic crop footprints to yield a probabilistic pesticide use footprint illustrating the probability that each pixel is both cropped and treated with the pesticide. The probabilistic pesticide use footprints can also be used to determine probabilities of pesticide occurrence in aquatic environments. To do so, the mean and maximum values of the probabilistic pesticide use footprint are calculated by subwatershed (Hydrologic Unit Code 12, the most detailed hydrologic units published by the US Geological Survey) and assigned to waterways within the subwatershed. Waterways and waterbodies that receive inflow from neighboring subwatersheds are also assigned the mean and maximum pesticide use probabilities of those larger watersheds to provide both conservative and more realistic estimates. All other waterbodies are assigned the values found within locally delineated watersheds. Probabilistic SDMs are constructed in APCOAT using species location records, environmental variables that may influence species distribution, and the maximum entropy machine learning method implemented by the Maxent software platform (Phillips et al., 2017). The advantages of using APCOAT to run the Maxent platform include the ability to model species in terrestrial and aquatic environments, the availability of prepared environmental predictor variables (e.g., biologically relevant climate averages, land cover, soil texture, stream flow rates), and iterative modeling designed to select the model with the best statistical fit that minimizes the number of included predictor variables. These features permit users to produce probabilistic SDMs by simply providing APCOAT with species location records, available through organizations such as the Global Biodiversity Information Facility, NatureServe, or iNaturalist. By modeling species distributions with these data sets and methods, users can visualize areas of high-quality habitat and potential restoration sites that may not be apparent from critical habitat maps, which may be unavailable for understudied organisms. In the final APCOAT analysis step, spatial co-occurrence between pesticide usage and species of interest is calculated by obtaining the product of their respective probabilistic models. The resulting co-occurrence model is then summarized within zones of interest, and a report is generated detailing the inputs, methods, and results of the co-occurrence analysis. These reports can then be used to rapidly identify organisms requiring higher-tier environmental risk assessments based on the average range-wide probability of co-occurrence or to target areas within the species geographic range for habitat conservation and outreach to pesticide applicators. By making the tools and data described herein publicly available and easily accessible, the barriers to rapidly and accurately assessing the potential spatial co-occurrence between pesticide use and species locations have been reduced significantly. Moreover, the inclusion of tools for aquatic species analyses offers the opportunity for significant expansion in understanding the distributions of the organisms, which have rarely been studied in this manner. APCOAT provides all stakeholders involved in endangered species assessments transparent access to scientific techniques necessary for accurately quantifying the spatial extent and likelihood of pesticide use that could potentially affect threatened and endangered species. Jonnie B. Dunne: Conceptualization; data curation; investigation; methodology; project administration; software; visualization; writing—original draft. Hendrik Rathjens: Conceptualization; investigation; methodology; software; supervision; visualization; writing—review and editing. Michael Winchell: Conceptualization; methodology; supervision; writing—review and editing. Max Feken: Funding acquisition; methodology; resources; supervision; writing—review and editing. Richard Brain: Funding acquisition; methodology; resources; supervision.