EVALUATING MULTI-YEAR TRANSFERABILITY OF SINGLE- AND MULTI-CLASS MACHINE LEARNING FOR REMOTE MONITORING OF BEAVER-ENGINEERED LANDSCAPES

Author

• Harris J. Sloan, The University Of MontanaFollow

Year of Award

2026

Document Type

Thesis

Degree Type

Master of Science (MS)

Degree Name

Ecology and Evolution

Department or School/College

Division of Biological and Biomedical Sciences

Committee Chair

Dr. Bret W. Tobalske

Commitee Members

Dr. Angela D. Luis, Dr. Robert O. Hall, Dr. Lisa A. Eby

Keywords

beavers, beaver dams, machine learning, satellite imagery, spatial ecology, remote sensing

Subject Categories

Biology

Abstract

Long-term satellite image archives enable ecological monitoring at large scales, but manual extraction of fine-grained features (e.g., beaver dams) remains time-consuming. Machine learning models may automate these tasks, but their temporal transferability for multiannual monitoring is still poorly resolved.

We evaluated the performance of two complementary neural-network frameworks across a 100,000-hectare watershed in Montana USA, using a decade of satellite imagery (2014-2024): (i) a single-class ResNet-50 beaver dam detector, and (ii) a multi-class ResU-Net landcover classifier.

In 2024, the dam detector identified 352 of 404 mapped dams, achieving balanced accuracy = 93.6%, precision = 79.8%, recall = 87.1%, F1 = 83.3%; spatial overlap between predicted and manually delineated dam footprints was low (IoU = 14.4%).

Across 2014-2023, dam detection performance remained generally high (𝑏𝑎𝑙𝑎𝑛𝑐𝑒𝑑 𝑎𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = 84.2%), but temporal variation was significant (𝑟𝑒𝑐𝑎𝑙𝑙 = 68.4%, 𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = 70.4%). The landcover classifier was accurate and temporally stable across 2014-2024 (𝑎𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = 86.2%, 𝑘 = 72%, 𝐼𝑜𝑈 = 72.7%), with strongest performance in water and forest classes and lowest performance in urban classes.

Synthesis and applications: single-class patch-classifiers can support multiannual monitoring of small, context-dependent ecological features such as beaver dams, but their sensitivity to interannual variation requires periodic manual validation and may benefit from temporally representative training data. Multi-class landcover classifiers provide stable, low-supervision landscape data that can be paired with single-class models to quantify ecosystem change and guide management decisions regarding beaver-mediated restoration, riparian condition, and watershed planning.

Recommended Citation

Sloan, Harris J., "EVALUATING MULTI-YEAR TRANSFERABILITY OF SINGLE- AND MULTI-CLASS MACHINE LEARNING FOR REMOTE MONITORING OF BEAVER-ENGINEERED LANDSCAPES" (2026). Graduate Student Theses, Dissertations, & Professional Papers. 12768.
https://scholarworks.umt.edu/etd/12768