This doctoral dissertation investigates the use of multi-sensor satellite image time series for forest monitoring in selected ecological regions of Slovenia. The research focuses on integrating optical (Sentinel-2, PlanetScope) and radar (Sentinel-1) data—specifically coherence products—for tree species mapping, remote sensing-based phenology analysis, and the detection of abrupt disturbances such as logging and natural tree mortality. The study covers the 2017–2021 period. The primary objective of the dissertation is to develop a methodology for annual tree species mapping using satellite image time series. The best classification accuracy (90%) is achieved with Sentinel-2 optical time series, followed closely by fused multi-sensor coherence and optical time series (89%). In contrast, using coherence time series alone results in a lower classification accuracy of approximately 75%. The classification outputs form the basis for analysing seasonal dynamics, remote sensing-derived phenology and abrupt forest disturbances, including logging and natural mortality events. The phenology results show that spectral indices based on red-edge, near-infrared (NIR) and shortwaveinfrared (SWIR) spectral bands—particularly IRECI, EVI, EVI2, SAVI, ARVI, and kNDVI—are the most informative for predicting key phenological stages, such as the start (SOS) and end (EOS) of the growing season. For detecting sudden disturbances, the NBSI and NDVI indices are particularly valuable as they most effectively capture vegetation degradation and changes in forest ecosystem condition. Bark beetle infestations cannot be reliably detected in their early stages using currently available satellite data (Sentinel-1, Sentinel-2, and PlanetScope). However, these data, when combined with in situ observations, can serve as a valuable training dataset for the future implementation of advanced approaches such as convolutional neural networks (CNNs), particularly when supplemented with additional data sources such as hyperspectral imagery.
|
