Plant phenology, the seasonal timing of recurring biological events, is a sensitive indicator of climate change and an important regulator of the biosphere. Currently, phenology is studied at three distinct spatial scales: (1) at the level of the individual using on-the-ground observations, (2) at the local landscape scale using elevated automated cameras called PhenoCams, and (3) at the regional scale using vegetation indices derived from satellite imagery. My research aims to bridge these scales so that seasonal knowledge gained about individuals can be translated across regions and continents, and so indices from PhenoCams and satellites can be more accurately biologically interpreted. In particular, the causes and consequences of climate change occur on multiple spatial and temporal scales, so that understanding and forecasting these changes requires integration across these scales. Ultimately, we want to be able to produce phenology forecasts to aid managers, policy makers, scientists, and the public in areas such as agriculture, forestry, public health, climate forecasting, and conservation. My two current projects in this area are:
1. Scaling up terrestrial plant phenology from individuals to continental scale
I am using data on phenology, abiotic conditions, and weather from the National Ecological Observatory Network (NEON), PhenoCams, satellites, long-term site-specific data sets, citizen science, and other sources to create a hierarchical Bayesian modeling framework for linking phenology measures across spatial scales. I am addressing the questions:
- How do local abiotic and biotic environments affect the phenology of individuals within species, across different plant functional types and regions?
- To what extent do combined ground-based phenology measurements of dominant species reflect community phenology measurements recorded by landscape-level sensors? Does the concordance vary by vegetation type or region?
- How can inferences about phenology drawn from satellite data be improved by incorporating fine-scale data about local abiotic and biotic environmental conditions and the responses of species to those conditions?
- To what extent can phenology measurements made at single sites be extrapolated to other regions? How does this extent depend on vegetation type or particular phenology drivers?
(Supported by NSF grant DEB-1550740)
2. Inferring plant phenology using PhenoCams together with citizen science, crowdsourcing, and machine learning
Andrew Richardson and colleagues have developed a technique for automatically extracting a measure of greenness from PhenoCam images. I am researching what additional information can be derived from these images – such as flowering, seed production, crop stage, and snow cover – using a citizen science project called Season Spotter that I developed on the Zooniverse platform. Additionally, I am using classifications from citizen science volunteers to ground-truth the automated greenness measures for spring and fall. I am also exploring the feasibility of using automated techniques on arbitrary public webcams that were not set up to be monitoring phenology (using the Archive of Many Outdoor Scenes developed by Robert Pless). Being able to use arbitrary webcams would expand the phenology camera network from hundreds of cameras primarily in North America to thousands around the world.
The existence of snow in PhenoCam images has complicated automated processing. I am developing a method to automatically flag images having snow by creating a dataset of 200,000 tagged images (using the Knowxel crowdsourcing platform) and using this dataset to train convolutional deep neural networks. This automated crowdsourcing-to-machine-learning approach has broad applicability to environmental image processing and may open new frontiers in automated environmental data collection.
(Supported by NSF grant EF-1065029)
African mammal community ecology
Intact large mammal systems allow us to better understand the spatial and temporal dynamics of natural systems at scales similar to those used by humans. These systems are also threatened by habitat transformation, poaching, and emerging diseases, making them a critical target of conservation efforts.
Multi-species long-term dynamics
My ongoing work focuses on the Serengeti ecosystem, stretching the border between Tanzania and Kenya. It contains many dozens of mammal species, including a diverse suite of large and medium-sized predators and almost 20 ungulate species. While individual species have been the target of long-term monitoring and research, long-term research investigating the interplay among multiple species at different trophic levels has been rare due to the complicated and expensive logistics.
In collaboration with Alexandra Swanson, Craig Packer, and the Zooniverse, I developed the wildly popular Snapshot Serengeti citizen science project that asks volunteers to identify animal species in images from a Serengeti camera trap network that Dr. Swanson designed and deployed and Dr. Packer maintains. The 200+ cameras produce millions of these images and human classification is currently the only way to produce accurate data from them. Using citizen science for image classification saves the project approximately $25,000 annually, while engaging tens of thousands of people in the scientific process.
Snapshot Serengeti data has provided insight into large predator coexistence. Current community ecology research includes predator-prey dynamics, insectivore coexistence, wildlife disease, and mammal migration. Additionally, Snapshot Serengeti has been used for research into wildlife modeling, generalizable citizen science techniques and computer vision, and has been used in collegiate classrooms as an integral part of introductory biology courses.
(Currently supported by the National Geographic Society)
Wildlife disease dynamics
Understanding the temporal dynamics of wildlife diseases requires knowledge of one or more hosts, the disease, and the environment in which the host(s) and disease live. Mathematical models can sometimes be used to predict dynamics, but frequently they are too overly simplified for real biological inference – especially when host(s) or diseases have non-linear spatial dynamics or there are feedbacks between disease and host behavior. I develop simulation and statistical models using a Bayesian approach that can take in as much information as is available and infer possible wildlife disease system parameters. I run these simulations on supercomputers, which allows for exploration of multi-dimensional parameter spaces in reasonable timeframes.
I have used an Approximate Bayesian Computation approach to infer the disease dynamics of bovine tuberculosis (a multi-host disease) in lions in Kruger National Park, South Africa. The results give insight into the important drivers of bovine tuberculosis in this system and provide predictions for the future. A full description of this research is provided in:
Kosmala, M., P. Miller, S. Ferreira, P. Funston, D. Keet, C. Packer. (2016). Estimating wildlife disease dynamics in complex systems using approximate Bayesian computation models. Ecological Applications, 26(1): 295-308.
Or, read the layman’s summary of this research.