Leveraging machine learning for maritime object detection and peatland classification : harnessing the power of machine learning for precise maritime object detection and peatland classification

Abstract

This thesis deals with two application sectors where Machine Learning (ML) approaches have a central role. The first one is the maritime environment, where one of the tasks is to create the Situational Awareness (SA) model, showing what is happening in the environment around a vehicle. The second sector focuses on peatland classification to characterize and differentiate various peatland types.

The maritime study proposed in this thesis investigates the most relevant target predictors in the maritime environment, focusing on different Convolutional Neural Network (CNN) architectures. Additionally, Transfer Learning (TL) is implemented to determine if its use enhances object recognition performance. Subsequently, a maritime dataset is developed. The dataset is precisely manually annotated and can be used for two main computer vision tasks: object detection and tracking. The purpose of the dataset is to provide a solid basis for the development of efficient MLbased approaches for SA modeling in maritime environments. Article I evaluates the performance of three state-of-the-art object detection algorithms using datasets collected in the Finnish archipelago. Best method is Faster R-CNN with ResNet101 as feature extractor achieving the highest accuracy at 74.0%. Article II addresses the limited availability of domain-specific datasets in maritime environments. For this purpose a new dataset was tested with various detectors, revealing that Faster R-CNN (35.18%) and EfficientDet (55.48%) achieved the highest average precision. Article III explores the performance of Faster R-CNN, R-FCN, and SSD using different feature extractors, with Faster R-CNN achieving the highest mean average precision at 75.2%.

The objective of the remote sensing part of the thesis is to create and evaluate a methodology that starts from a set of Geographic Information System (GIS) data input and finishes with the output of a soil-type classification map, especially focusing on pixel-wise soil-type classification. The proposed peatland methodologies summarizes the accumulation of decay material. Peatland areas are mainly found where vegetation decomposition exists. Peatland areas help to regulate the vegetation state and water availability. Article IV proposes a CNN fusion approach for peatland site type classification by integrating multi-source and multi-resolution data, achieving an accuracy of approximately 32%. Article V investigates the performance of CNNs when trained with a high number of synthetic aperture radar (SAR) and visual bands (51.06%) compared when trained with only the best bands (56.73%). Article VI extends the methods used in Articles IV to different zones in Finland, achieving a classification accuracy ranging from 26.9% to 33.6%.