Using Stochastic Processes and Machine Learning to Solve Environmental Challenges: Applications in renewable energy and remote sensing

This thesis investigates the application of artificial intelligence (AI), machine learning, and environmental statistics to address challenges in renewable energy optimization and environmental monitoring. Environmental systems are inherently complex, dynamic, and uncertain, making data-driven methods particularly valuable for modeling and analysis. Advances in remote sensing technologies and the increasing availability of large-scale environmental data have further enabled the development of AI-based approaches capable of extracting spatial and temporal patterns from heterogeneous and incomplete datasets.

The first part of this thesis focuses on the optimization of Wave Energy Converters (WECs). Ocean waves represent a promising renewable energy source due to their high energy density and predictability, but efficient energy extraction requires adaptive system design under stochastic sea conditions. Using statistical models of ocean wave behavior, Papers I and II investigate the optimization of a WEC system developed by the company Waves4Power.% In Paper~I, a simplified model is used to optimize accumulator pressure in the power take-off (PTO) system based on sea state characteristics. Paper II extends this work using a more detailed system model to improve generator performance and operational efficiency. The results demonstrate that environmental statistics can effectively guide adaptive control strategies and improve WEC robustness under varying ocean conditions.

The second part of the thesis addresses environmental monitoring using aerial and satellite imagery. In Paper III, the focus is on the early detection of bark beetle infestations in forests using Sentinel-2 satellite time series data and static environmental features classified by XGBoost.

In Paper IV, a Conditional Variational Autoencoder (CVAE) combined with a Gaussian Mixture Model (GMM) is proposed for semi-supervised classification by integrating temporal satellite observations with static environmental variables.

In Paper V, a U-Net-based deep learning model is developed to reconstruct distorted or noisy Shortwave Infrared (SWIR) spectral information in optical remote sensing imagery.

Although the applications considered in this thesis differ, both are unified by a common methodological perspective: the use of data-driven models to analyze and optimize environmental systems characterized by uncertainty, temporal variability, and incomplete observations.