Aidan Wright

Topic

A Machine Learning Approach for the Detection of Kelp-Encrusting Bryozoans in UAV-Derived Low Altitude Imagery

Department of Geography

Date & location

• Friday, September 12, 2025

• 9:30 A.M.

• David Turpin Building

• Room B215

Reviewers

Supervisory Committee

• Dr. Maycira Costa, Department of Geography, University of Victoria (Co-Supervisor)

• Dr. Neil Ernst, Department of Computer Science, UVic (Co-Supervisor) 

External Examiner

• Dr. Henry Houskeeper, Applied Ocean Physics & Engineering, Woods Hole Oceanography 

Chair of Oral Examination

• Dr. Charles Curry, School of Earth and Ocean Sciences, UVic

   

Abstract

The ecological, cultural and economic importance of canopy-forming laminarian kelp has prompted widespread monitoring efforts in recent years. Some of these efforts have capitalized on low-cost uncrewed aerial vehicles (UAVs) equipped with RGB CMOS and multispectral (MS) sensors to assess kelp bed extent, spatio-temporal distribution and overall resilience in the face of a changing climate. The goal of this research was to develop a UAV-based methodology for detecting epiphytic Kelp-Encrusting Bryozoans (Membranipora membranacea, or “KEB”) on kelp. To achieve this, two objectives were executed. First, several Machine Learning (ML) classifiers were evaluated on their ability to detect naturally occurring KEB in ultra-high spatial resolution (~ 2 mm²) imagery using multispectral (G, R, RE, and NIR) and visible RGB CMOS-derived bands. This UAV-derived imagery was acquired in the spring and summer of 2024. Second, the effects of environmental conditions on KEB detection accuracy were quantified to recommend thresholds on acceptable conditions for image collection. The ultra-low altitude (~ 5 m) imaging protocol employed here was effective at acquiring imagery of sufficient quality to be used for the detection of KEB. Spectral confusion between KEB and other marine classes (like glint and floating debris) was high in most MS bands and indices according to Jeffries-Matusita and Transformed Discriminant scores (TD < 1.5, JM < 1.5). However, KEB-pixel reflectance values were statistically distinct from all other classes in at least one band (p_adj < 0.004). Of the ML algorithms tested, the XGBoost model had the highest KEB classification performance (F₁ = 0.852) of any model trained from the CMOS-derived RGB bands, MS bands, and vegetation indices for each pixel. A single hidden-layer MLP Neural Network was the best performing model out of those exposed only to the RGB bands (F₁ = 0.793) of the same training pixels. These two models were both highly accurate at distinguishing KEB from water, kelp, and glint (> 95 %). The MS-XGBoost performed 13.2 % to 8.9 % better than the RGB-NNet at distinguishing between spectrally similar KEB-debris and KEB-epiphyte class-pairs. Variable importance (VImp) scores revealed that the predictors associated with the largest reduction in classification error were the three CMOS-derived RGB bands rather than any of the MS bands or indices. A practical detection test evaluated the models' ability to classify imagery based on KEB “presence” or “absence”, using a KEB coverage detection threshold. Both RGB-NNet (TSS = 0.394, AUC = 0.699) and MS-XGBoost (TSS = 0.464, AUC = 0.707) models correctly identified most images that contained KEB (10 out of 12 and 11 out of 12) respectively. However, they did so with very low Precision (< 8 %) also categorizing many KEB-free images as “presence”. Using a cluster analysis, KEB pixel labelling accuracy was found to statistically decrease for images captured in unfavourable environmental conditions (0.01 < p_adj < 0.023) such as high Beaufort category, high solar altitude, and mixed cloud cover. KEB false positives were minimized in classifications when images were collected with (1) a Beaufort category of 0; (2) a solar altitude below 40°; and (3) a cloud cover of either 90–100% or 0–10%.