Graduate Defense: Louis Jochems

Dissertation Information

Title: Tracking the Expansion of an Invasive Plant in Coastal Wetlands: Approaches Using Detection, Prediction, and Inference

Program: Doctor of Philosophy in Ecology, Evolution, and Behavior

Advisor: Dr. Jodi Brandt, Geosciences and Biological Sciences

Committee Members: Dr. Trevor Caughlin, Biological Sciences; Dr. Megan Cattau, College of Innovation and Design; Dr. Matt Williamson, College of Innovation and Design; and Dr. Kelly Hopping, College of Innovation and Design

Abstract

Intensifying land use, climate change, and human transport of species increase the susceptibility of ecosystems to novel conditions and species invasions. One example of these processes at play are the Great Lakes coastal wetlands. Disturbances from agricultural and suburban development in the region have led to the subsequent extirpation of native plant communities by dominant stands of invasive plants. Thus the ability to predict spread into uninvaded yet vulnerable wetlands remains a critical task to prevent further harm to these ecosystems. One invasive species, the floating-leaf aquatic plant, Hydrocharis morsus-ranae (European frogbit; hereafter EFB), has spread throughout coastal wetlands in Michigan, USA but with limited establishment at a few sites. Given its current expansion, one effective strategy to mitigate new invasions is through early detection and rapid response protocols (EDRR). The availability of remote sensing (RS) and field observations can improve these protocols via species distribution models (SDMs). For my dissertation, I used various data sources and SDMs to detect, predict, and gain inference on European frogbit’s distribution, all of which have implications for EDRR of this species.

For chapter 1, I focused on EFB detection via high-resolution RS imagery, since detecting newly established invasive plants is key to preventing further spread. Traditional field surveys are often insufficient to identify the presence and extent of invasions. This is particularly true for wetlands because of difficult access, and because floating and submergent plants may go undetected in the understory of emergent plants. Unoccupied aerial systems (UAS) are revolutionizing how we monitor invasive vegetation in wetlands, but key components of the analysis workflow have not been defined. In this study, I compared different methodologies for mapping Emergent, Floating (EFB and other species), and Submergent vegetation using a machine learning classifier on UAS imagery of a Great Lakes wetland. I compared predictive accuracies using (a) different spatial resolutions (11 cm vs. 3 cm pixels), (b) two classification approaches (pixel- vs. object-based), and (c) including structural measurements. The 11 cm imagery yielded higher accuracy metrics than the best-performing models of the finer (3 cm) resolution data. At each resolution, the top-performing models were from pixel-based approaches and included structural data over those with multispectral data alone. Overall, high-resolution maps from UAS classifications will improve the EDRR of invasive plants in coastal wetlands throughout the globe.

For chapter 2, I investigated how dispersal-related and ecological drivers affect the introduction and establishment of EFB in a spatially-explicit Bayesian SDM. Since EFB has not yet filled its niche in Michigan, dispersal dynamics such as proximity to known presences (spatial autocorrelation) and/or human vectors, may control its spread as much as habitat suitability. I quantified the importance of ecological, and dispersal-related drivers on the introduction (presence and absence) and establishment (abundance in percent cover) of EFB. I fit Integrated Nested Laplace Approximation (INLA) hurdle models to 11 years of field observations from coastal wetlands in Michigan and produced posterior distributions of model parameters and spatial random effects. EFB occurrence was most strongly and nonlinearly associated with water depth (cm), demonstrating that a specific ecological niche is an important limitation on EFB spread. EFB occurrence was negatively associated with distance to the nearest public boat launch, indicating that human recreational vectors are an important spreader of EFB. Finally, EFB was positively associated with cattail (Typha spp.) cover, which we attribute to the emergent functional type of cattails protecting EFB from wave energy of the Great Lakes. None of the included predictors had any meaningful effect on EFB abundance, suggesting that I either did not include important factors for establishment, or it may be too early in EFB’s invasion stage to fully quantify important drivers for this process. The proximity to known invasions was also an important for EFB spread since including a spatial random effects term improved model fit. These results indicate that incorporating spatial autocorrelation and dispersal factors into SDMs of an invasive species improves our ability to anticipate new occurrences of an invasive species that has not yet filled its niche. However, additional predictors over time or different spatial scales may be necessary to identify the drivers of this invasive species’ abundance. This framework informs EDRR of invasive species by providing quantified estimates (and their uncertainties) of important dispersal-related and ecological drivers.

For chapter 3, I assessed the predictive capacity of machine-learning SDMs based on remote sensing and geospatial datasets that proxy dispersal and habitat conditions for EFB during a period with climatic extremes. Specifically, I investigated whether two active remote sensing datasets, bathymetry from aerial light detection & ranging (LIDAR) and satellite-based synthetic aperture radar (SAR), can proxy water depth and emergent vegetation, respectively, as two key drivers of EFB. I compiled 1,100 occurrence records with eight remote sensing (both active and passive) and geospatial layers as predictors to train random forest models and generate performance metrics for predicting habitat suitability in coastal wetlands. Moreover, I accounted for climatic extremes (water levels) during the study period to investigate how predictions of EFB distribution may change between years. I found that the model containing all remote sensing and geospatial predictors yielded a mean overall accuracy of 84.5 % for predicting EFB occurrence. Active remote sensing data proxied emergent vegetation and water depth based on field measurements, but were relatively less important for predicting EFB than other ecological drivers. The area of highly suitable habitat (pixels assigned > 60% probability) varied considerably between the years of highest and lowest water levels during the study period. Despite variations, the influence of EFB’s dispersal dynamics (boat launches) outweighed other biophysical components for establishment during this early invasion stage. Overall, our findings suggest that integrating different types of remote sensing and geospatial data is required to predict potential EFB distribution across dynamic coastal wetlands.

In summary, these results provide insight into the ecological and dispersal dynamics of an invasive species’ distribution throughout Michigan’s coastal wetlands. Assessing the strengths and limitations of SDMs for predicting spread will become increasingly important for EDRR as climate change exacerbates the impact of invasive species on ecosystems.