iNaturalist mammal observations classified by evidence type

Motivation. Ecologists show growing interest in observational data generated by participatory scientists. For mammals, the largest participatory science platform is iNaturalist, which has more than 2.8 million observations from North American represented through images of living animals, dead animals, tracks, and scat. These different types of evidence could give insight into the underlying sampling paradigm for an observation (e.g. dead animals might be more likely to be reported near roads) and thus may be useful for scientific applications of these data. However, while iNaturalist allows users to annotate observations by evidence type, many observations are not annotated. We use machine learning (ML) to classify the evidence types associated with observations of North American mammals in iNaturalist, adding metadata that can be used to subset data or to model multiple observation processes. Main types of variable contained. This dataset contains metadata augmenting 1.33 million North American mammal iNaturalist observations with evidence type. Each observation is categorized as either live animal, dead animal, tracks, scat, or other sign, and an associated confidence score is provided. Spatial location and grain. North America. Each data point is associated with a single point in space. Coordinates and positional accuracy for each observation are provided. Time period and grain. Observations occurred between December 1, 1948 and February 27, 2024, although 99.89% of observations occurred since the year 2000. Each data point is associated with a single point in time. Major taxa and level of measurement. All North American species in the class Mammalia, excluding species in the order Chiroptera (bats). Observations are identified to the species level and meet iNaturalist’s requirements for “research-grade” data quality. Software environment. Data are supplied as comma-separated values (.csv) files. We also provide the trained ML model in (a) an interactive online portal and (b) a set of weights that can be used to reproduce predictions or fine-tune the model on new datasets using the provided source code. Note: v0.0.2 corrects an error where some rows had NA for year/month columns.