Remote sensing and machine learning-based solutions for ecosystem management and the provision of large-scale and long-term vegetation information

Global change is deeply connected to forest decline and biodiversity loss. To map and monitor these processes across spatial and temporal scales, vegetation information obtained from spaceborne remote sensing time series data is highly beneficial. The central objective of this thesis is to support national agricultural and forest agencies in their duties for providing large-scale, up-to-date and area-wide information to stakeholders and decision makers. This research investigates 1) the use of spaceborne remote sensing data for biodiversity and ecosystem mapping; 2) the development of data sampling strategies from real-world reference data for improved spatial predictive modeling and 3) the provision of datasets, code and models allowing practitioners to take action. The cumulative dissertation is based on three third-party funded research projects. The first project “Riparian forests in arid regions of Central Asia under pressure by use and limited water resources: definition of tipping points to ensure sustainable management” (ForeCeA) explored the remaining desert poplar forest stands (Populus euphratica, P. pruinosa) in Central Asian floodplains along the Syr Darya and the Ili river. The second project “Entwicklung eines automatisierten Hinweissystems für die Kontrolle von Agrarförderflächen” (timeStamp) investigated the spatio-temporal prediction of the EU’s common agricultural policy (CAP) greening measures in Germany based on the example of winter catch crop cultivation. The third project “Künstliche Intelligenz mit Erdbeobachtungs- und Multi-Source Geodaten für das Infrastruktur-, Naturschutz- und Waldmonitoring” (TreeSatAI) explored artificial intelligence (AI)-based methods for improved forest mapping and tree species classification. The thesis is divided into six chapters. Chapter 1 introduces the background, motivation and challenges of spaceborne remote sensing and AI for administrative vegetation mapping and monitoring. It ends with the research questions and objectives of this thesis. Chapter 2 aims at developing methods to map and monitor the remaining riparian forest stands at Central Asian floodplains with low- to medium-resolved satellite image time series from the MODIS and Landsat programs. Chapter 3 aims at the support of nationwide controls of winter catch crop cultivation in Germany with Copernicus data and predictive modeling. Chapter 4 proposes guidelines for the application of machine learning and deep learning techniques in forest remote sensing using the TreeSatAI Benchmark Archive, which contains multi-source and multi-labeled image patches from 50,000 forest stands in Northern Germany. Chapter 5 investigates the spectral and temporal content of spaceborne polarimetric and optical data in Central European temperate forests to provide recommendations for future tree species classification. Chapter 6 summarizes the main findings and describes remaining challenges, possible solutions and implications for future research.