Machine Learning for Classifying Marine Vegetation from Hyperspectral Drone Data in the Norwegian coast

Abstract

Along the Norwegian coasts the presence of blue forests are the key marine habitats. Due to increased anthropogenic activity and climate change, the health and extent of the blue forests is threatened. However, no low-cost, reliable system for monitoring blue forests exists in Norway at this time. This thesis studied machine learning methods to classify marine vegetation from hyperspectral data acquired in Norway. The study area is situated by Larvik at Ølbergholmen. The dataset consists of 12 hyperspectral images with 173 spectral bands in the region 390 nm - 749 nm and corresponding labels of the different classes. This dataset was used to train and evaluate the machine learning methods. In addition, an independent dataset from a different site was used for robustness evaluation. Three machine learning methods were studied; Random Forest (RF), Support Vector Machines (SVM) and Convolutional Neural Network (CNN). The results indicate that the powerful CNN approach had the best performance during validation based on the computed statistical measures. However, when evaluated for robustness, RF performed the best. The computed confusion matrices for the validation and robustness studies revealed that the presence of a so-called turf algae caused difficulties in distinguishing between the classes, which is an important finding with regard to future research. This thesis has shown that machine learning can be used for monitoring blue forests and various marine vegetation species using hyperspectral drone imaging along the Norwegian coast.