Archived

This content is available here for research, reference, and/or recordkeeping.

Author ORCID Identifier

https://orcid.org/0009-0002-7405-7694

Date Available

5-8-2027

Year of Publication

2026

Document Type

Master's Thesis

Degree Name

Master of Arts (MA)

College

Arts and Sciences

Department/School/Program

Geography

Faculty

Liang Liang

Faculty

Nick Lally

Abstract

Understanding species-level phenology is crucial for assessing ecological responses to climate and environmental stress, especially in urban areas where microclimate variability and vegetation heterogeneity create complex growth patterns. While conventional remote sensing has enabled large-scale vegetation phenological studies, its coarse resolution and limited spectral detail hinder the detection of fine-scale, species-specific phenological dynamics in urban landscapes. This study investigates the potential of multi-index PlanetScope phenometrics, time-series features, and seasonality data for phenological characterization and machine learning classification of 18 urban tree species across the University of Kentucky main campus in Lexington, Kentucky. Six phenophases were extracted from a time series of four vegetation indices, NDVI, EVI2, CIRE, and NDRE. The phenophases were derived using the amplitude threshold method for more than 2600 trees. Kruskal-Wallis tests confirmed statistically significant inter-species phenological differences across all 96 index-phenophase-year combinations (p< 0.001). The mid-senescence phase produced the highest effect size for NDVI (ε² = 0.211), and CIRE produced the highest effect size for MidGreenup (ε² = 0.203). A consistent pattern of complementarity emerged as greenness-based indices were superior for species discrimination during senescence phase and red-edge indices were superior for green-up discrimination. Inter-annual rank consistency confirmed that Mid Senescence rankings were the most stable across the three years of the study period (r=0.77 to 0.88). Random Forest classifier consistently outperformed XGBoost across all four indices. The best model (NDVI all features) achieved balanced accuracy of 77.5% and overall accuracy of 56.1% across 18 species. Species with more distinct phenological properties identified in the statistical analysis were classified more accurately. In comparison, the species with similar phenological properties had poor classification accuracy. Overall, this demonstrates that the approach used in the study can contribute to developing a scalable method for high-resolution vegetation monitoring.

Digital Object Identifier (DOI)

https://doi.org/10.13023/etd.2026.248

Archival?

Archival

Funding Information

The study was supported by Barnhardt-Withington-Block (BWB) Funding – Summer Funding

Download

Available for download on Saturday, May 08, 2027

Contact Author

Share

COinS