Pixel-Wise Classification of Live and Dead Maerl Using Unsupervised and Semi-Supervised Learning

Maerl beds—formed by calcareous red algae—are ecologically valuable habitats that provide structure, biodiversity, and ecosystem services in coastal environments. Monitoring their condition requires distinguishing between live and dead maerl, a task typically performed through manual interpretation of underwater imagery. This process is time-consuming, subjective, and difficult to scale. In this thesis, a combined unsupervised and semi-supervised learning pipeline is proposed to automate pixel-wise segmentation of live and dead maerl in high-resolution images. The pipeline begins with feature-based clustering using Gaussian Mixture Models (GMMs) and the Expectation-Maximization (EM) algorithm, applied to selected color and texture features. This enables the generation of initial segmentation masks without manual labels. A semi-supervised approach further improves segmentation by incorporating weak supervision in the form of premasks and estimated class proportions.These masks are then used to train a DeepLabv3 semantic segmentation model with a ResNet-50 backbone. Iterative self-training and manual refinement result in a final dataset of 80 annotated images, tiled into over 5,000 patches fortraining and evaluation. The DeepLabv3 model achieved a Dice coefficient of 0.901 and a pixel accuracy of 60.3% on the test set. While the model effectively segments regions with strong visual cues, challenges remain in classifying ambiguous boundaries and rare cases. The results demonstrate that combining unsupervised clustering with CNN-based segmentation offers a scalable, and more efficient solution for habitat monitoring. The approach significantly reduces manual annotation effort and provides a frameworkfor future deployment in maerl habitat monitoring.