128-4 A Methods-Based Framework for Assessing Riparian Vegetation Health in Human-Impacted and Natural Streams Using UAS Imagery

Abstract:

Riparian vegetation is important for river health because it reduces erosion, controls runoff, regulates water temperature, and stabilizes streams. However, the southwestern United States has lost an estimated 80–95% of its riparian vegetation due to human activities such as land-use change, mining, and deforestation. Monitoring remaining vegetation is important for understanding how streams respond to disturbance, but traditional field methods can be time-consuming and limited in area. Unmanned aerial systems (UAS) provide high-resolution imagery over larger areas in less time.  This study develops and evaluates a UAS-based machine-learning approach for mapping riparian land cover at a human-impacted stream and a natural reference stream in central Kentucky.

RGB imagery was collected at Camden Creek, a human-impacted stream, and Cane Run, a natural reference stream, and processed into high-resolution orthomosaics. Trees, grass, riparian grass, algae, water, sediment, shadows, and human-made features were manually outlined to create training data. Random Forest Classification (RFC) was used to classify land cover at the pixel level, with training data refined to improve classification. Segment Anything Model 2 (SAM2) was used to identify feature boundaries from image tiles using reference polygons and point prompts. SAM2 performance was measured using Intersection over Union (IoU), which compares predicted and manually outlined features. The Visible Atmospherically Resistant Index (VARI) was also used to compare vegetation greenness.

Results show that refining training data improved RFC class separation, while SAM2 successfully identified feature boundaries. Trees had the highest SAM2 performance (mean IoU = 0.917), followed by algae (0.855) and riparian grass (0.841). Visually similar features were more difficult to separate, likely due to similar colors, changing light conditions, and shadows. Combining SAM2 with RFC reduced pixel-level noise and produced cleaner feature boundaries, changing 14.7% of classified pixels. Similar performance between RFC and the combined classification suggests that SAM2 improved feature boundaries more than land-cover identification. VARI also showed differences in vegetation greenness, with riparian grass having the highest values.

These results show that combining UAS imagery with two machine-learning approaches can improve riparian land-cover mapping. However, RGB imagery alone may not be enough to separate vegetation types with similar colors and appearances. This approach provides a framework for future monitoring of riparian vegetation and changes associated with human disturbance, erosion, flooding, and channel change.

© Copyright 2026 The Geological Society of America (GSA), all rights reserved.