Spatiotemporal Characterization of Land Cover with Remote Sensing: Exploring Training Data Strategies and Sentinel-2 Time Series

Abstract

This thesis explores two key areas at the intersection of land cover and remote sensing: land cover mapping and continuous monitoring. While using machine learning for land cover mapping has become widespread, a notable gap in the literature exists regarding the organization and synthesis of training data approaches. Furthermore, there is a lack of studies exploring methods to reduce the heavy dependence of state-of-the-art techniques, such as deep learning, on training data. On the continuous monitoring side, although several studies have used Landsat or harmonized Landsat-Sentinel-2 data, there is a limited focus on utilizing Sentinel-2 data alone for continuous land cover monitoring. To address these gaps, this thesis proposes a series of studies to answer specific research questions related to land cover mapping and monitoring. A systematic review is conducted to synthesize how training data has been addressed in the remote sensing literature. Existing point-based reference land cover data is integrated with weakly and self-supervised learning for national scale land cover mapping in Portugal. Lastly, the feasibility of using Sentinel-2 data for continuous forest loss monitoring with the Continuous Change Detection and Classification algorithm is assessed in a dynamic landscape in Portugal. The research results include a comprehensive synthesis of the various methods used for training data in land cover mapping, providing a valuable guide for future research. Additionally, the thesis demonstrates the effectiveness of combining deep learning with weakly and self-supervised learning and leveraging reference datasets to reduce the need for fully annotated training data. Finally, the research shows that Sentinel-2 time series can enable accurate, spatially detailed, and agile monitoring of forest loss, showcasing its potential for continuous land cover monitoring. These findings contribute to advance the knowledge in the field by opening new avenues for more efficient and scalable land cover mapping and monitoring.