Central European forests provide many ecosystem services such as climate regulation, carbon storage, and economic revenue. However, they are increasingly threatened by forest disturbances driven by climate change. Forest disturbances in Central Europe arise from a variety of agents, including logging, windthrow, drought, and bark beetle infestations, which exhibit highly heterogeneous spatial and temporal dynamics. They are often small, subtle, and develop gradually over time. Bark beetle infestations, for instance, are a key forest management challenge, as they need to be detected within a time frame of 10 weeks after infestation to prevent further dispersal, while visible infestation symptoms usually occur later. To support effective forest management, continuous large-scale monitoring is required. A suitable means for that is multispectral satellite remote sensing for its sensitivity to physiological properties of forest canopies. The Sentinel-2 (S2) satellite constellation provides freely available observations with frequent revisits and moderate spatial resolution. This results in dense satellite time series reflecting forest dynamics through temporal patterns of spectral signals. To analyse these patterns, two monitoring paradigms can be distinguished: offline and online monitoring. The former is used to retrospectively detect disturbances for strategic planning, whereas the latter focuses on near-real-time detection of the onset of disturbances for tactical forest management responses. Both monitoring paradigms require the analysis of noisy, irregular, and complex satellite time series data, where disturbance signals are often subtle and partly overlapped by environmental variability. Conventional approaches based on predefined temporal features or statistical thresholding often struggle to capture such complex and nonlinear patterns in a scalable and transferable manner. On the contrary, the growing S2 archives enable Big Data approaches such as Deep Learning (DL), which can learn complex nonlinear temporal patterns directly from large and diverse time series datasets. This allows them to capture subtle and nonlinear disturbance signals that are difficult to model with conventional approaches. Therefore, this thesis aims at advancing the development of DL approaches for offline and online forest disturbance monitoring that are scalable across large areas, with a focus on the detection of small and subtle forest disturbances and the assessment of their detectability limits. The thesis is based on three studies that focus on Germany and Luxembourg as representative Central European countries: The first study presents a DL-based offline forest disturbance monitoring method using irregular S2 time series trained on a large and diverse dataset across heterogeneous forest ecosystems in Central Europe. The model is specifically trained to focus on small disturbances. Through a spatial hold-out validation design, the generalizability to unseen forest ecosystems including different disturbance agents is evaluated. The model achieves state-of-the-art performance even when taking sub-pixel-level disturbances into account. Nevertheless, scalability limits become apparent when ecological conditions change significantly, i.e., under ecological domain shift conditions. The second study develops one of the first DL-based online monitoring methods for reconstruction-based anomaly detection using irregular S2 time series in Central Europe in an operational setting. The model is trained on a large and diverse dataset, partly derived from automated high-resolution disturbance products, enabling scalable data generation. It is evaluated on a dataset with precisely known bark beetle infestation dates under a spatial hold-out validation design, allowing assessment of detection delay under realistic operational conditions. While reliable detections are achieved after approximately 13 weeks, the model fails to consistently detect infestations within the operationally required 10-week detection window, indicating challenges in early detection despite strong overall performance and transferability. The third study benchmarks conventional time-series-based online monitoring methods against the DL-based approach from Study 2 for timely detection of bark beetle infestations. Under realistic operational conditions, including spatial hold-out validation and evaluation of detection date instead of retrospective breakpoint assessment, the DL method clearly outperforms conventional approaches, primarily by reducing false positives. However, all methods fail to reliably detect infestations within the critical early detection window. This indicates limitations that are unlikely to depend solely on the applied algorithms. Taken together, the three studies provide evidence on the capabilities and limitations of DL-based forest disturbance monitoring using S2 time series under realistic large-scale conditions. This thesis advances large-scale forest disturbance monitoring capabilities in Central Europe by developing DL models for offline and online monitoring trained on heterogeneous datasets using irregular S2 time series with robust performance, including production pipelines and – in case of Study 1 – operational deployment. A spatial hold-out validation was consistently carried out to support understanding large-scale application capabilities and potential performance constraints under ecological domain shift in realistic operational settings. While good generalization of the developed DL models was confirmed throughout the studies, a moderate performance decline could be noted in ecological domain shift conditions in Study 1, i.e. when ecological conditions strongly changed, particularly in case of small disturbances. This indicates ecological information constraints when scaling DL models to larger areas. Increasing ecological variability leads to stronger class overlap, making small and subtle disturbances particularly difficult to detect. I conclude that scaling DL models to large areas across heterogeneous ecosystems leads to informational limits of the spectral signal. Furthermore, the inability to detect bark beetle infestations within the critical early detection window throughout modelling paradigms including DL indicates that this limitation is likely not solely methodological, but may reflect fundamental signal constraints of S2 time series with respect to early bark beetle infestation detection in Central Europe. To control the impact of ecological variability on spectral variability, future studies must incorporate ecological anchors such as auxiliary environmental variables as model input to reduce intra-class variability. This would enhance class separability across large areas, leading to improved large-scale applicability of DL methods.
