COS 311-5 - Exploiting the microclimate by-catch of wildlife cameras

Abstract: Microclimate – proximal climatic variation at scales of meters and minutes – is a crucial mediator of the impacts of climate change on wildlife. However, microclimatic data are hitherto temperature-centric, while factors such as sunshine, rare meteorological events (e.g. hail) and snow cover are rarely recorded at high resolution. Meanwhile, the use of cameras to record wild plants and animals is booming, as are deep learning models to extract ecological data from images. Wildlife cameras generate rich datasets about species’ behaviors, life cycles, as well as intra- and interspecific interactions. Here we postulate that they also generate a wealth of latent climatic information to explain those phenomena.

Here we leverage microclimatic by-catch of time-lapse cameras recording plants and small animals in two mountain regions: The Swiss Alps and the South African Maloti-Drakensberg. We present an approach to generate vast and comparable microclimatic datasets, covering both sites, in four steps: First, we annotate qualitative microclimatic conditions in a subset of 8,953 images spanning 49 plots. Second, we train and test ensemble deep learning models to detect evidence of sunshine, snow, and hail. Third, we classify and analyze microclimatic conditions in almost 2 million unannotated images. Fourth, a set of production models is tested on out-of-distribution images from Norway.

Ensemble models consistently achieve >90% out-of-sample accuracy. Furthermore, disagreement between ensemble member models forms a useful measure of uncertainty; by discarding non-consensus predictions, we achieve an effective accuracy >95%. Remarkably, when testing our models on images from a site in Norway, >1,000km from our training sites, our models distinguish sunshine from overcast conditions with up to 99% accuracy. Such an uncommon out-of-distribution performance suggests this approach has great potential.

The resulting data are extremely useful to explore spatial and temporal aspects of microclimate. We demonstrate consistent patterns of summer hailstorms and autumn snow across mountain sites in the northern and southern hemispheres. Sunshine and cloud cover clearly drive diel temperature cycles, while the South African site was found to be characterized by sunny mornings and cloudy afternoons. We demonstrate application of these microclimatic data by exploring (1) the impact of warming treatments on the persistence of autumn snow, and (2) the microclimatic foraging preferences of Bombus bees at high elevation. Our methods and models will allow a burgeoning community of camera trap and phenocam users to cheaply quantify fine-scale microclimatic variation, and to understand the microclimatic drivers of the ecological phenomena they record.