Author ORCID Identifier

https://orcid.org/0000-0002-2795-5996

Date of Award

Summer 7-20-2026

Document Type

Thesis (Ph.D.)

Department or Program

Ecology, Evolution, Environment and Society

First Advisor

Matthew P. Ayres

Second Advisor

Hannah M. ter Hofstede

Abstract

Ecological inference about bird communities depends on ecological processes that structure communities and the observation systems used to measure them. This dissertation integrates these dimensions across forests of the northeastern United States by examining how forest conditions shape bird communities and how biodiversity-sensing technologies influence ecological inference. Advances in sensors, autonomous monitoring, and artificial intelligence are expanding the spatial and temporal scales of ecological observation. However, AI-assisted observations introduce new error structures, making calibration, validation, and treatment of uncertainty essential for reliable inference.

First, I tested how forest productivity, composition, maturity, vertical structure, and invertebrate prey jointly affected species richness and community bird energy, a metabolically scaled measure of assemblage-level energy use. Forest greenness was strongly associated with invertebrate availability, but both avian metrics responded nonlinearly, peaking at intermediate greenness. Community bird energy also increased with forest maturity and vertical canopy complexity, while direct relationships with measured prey biomass were weak. These findings indicate that regional species–energy relationships emerge from interacting energetic, structural, compositional, and trophic processes.

I then evaluated passive acoustic monitoring combined with the deep-learning classifier BirdNET. I developed and independently tested a logistic-regression framework for estimating species-specific precision thresholds from manually validated detections. Fixed universal thresholds produced highly variable precision among species, whereas species-specific thresholds improved consistency and reduced false positives, although higher precision targets reduced recall. Independent evaluation showed that thresholding was effective only when score distributions of correct and incorrect detections were sufficiently separated.

Next, I compared ecological inference drawn about avian communities from five observational workflows including field-based and AI-assisted acoustic observations. Relative spatial patterns in richness were broadly preserved, but absolute richness, occupancy, detection estimates, and uncertainty differed among workflows. Extended temporal sampling improved occupancy precision. Finally, I produced approximately 217 hours of strongly annotated forest soundscapes as an open resource for classifier training, benchmarking, and ecological research.

Together, this dissertation shows that observation workflow is an integral component of ecological study design. Robust AI-assisted biodiversity monitoring requires species-specific calibration, independent validation, representative reference datasets, transparent processing decisions, and statistical models matched to the information content and error structure of acoustic data.

Available for download on Monday, July 31, 2028

Share

COinS