WP4 Fernerkundung und maschinelles Lernen (externer Projektpartner Prof. Miguel Mahecha von der Universität Leipzig)

Objective: Satellite remote sensing is one of the most consolidated tools for monitoring forests over large areas and long-time spans. Today, the European Sentinel sensors provide an unprecedented amount of data with excellent radiometric resolution in space and time, and spatial coverage. Furthermore, the availability of additional spectral bands, like the “red-edge”, opens new possibilities to monitor forest health and productivity. Even though these new sensors have high potential to improve the monitoring of forest stress, their short operational period of only a few years implies that practical experience is still limited. The objective of this WP is to explore the use of the “red-edge” bands of the optical domain (Sentinel 2) to monitor forest health and of radar data (Sentinel 1) to detect structural changes in forests. In addition, it is planned to explore the potential of high spatial resolution sensors (e.g. RapidEye), which come with less spectral information but have longer temporal coverage, to monitor forest stress. The overarching objective is to use modern AI methods to translate data streams into interpretable products by jointly exploring remote sensing and in-situ data.

Approach: We will develop a processing chain that automatizes the satellite data download and processing into “Analysis Ready Data Cubes” (Mahecha et al. 2020). Specifically, we will provide data in cloud-ready formats (zarr), so that they can be visualized by all project partners and continuously updated. The WP will then focus on fusing and co-interpreting multiple data streams with state-of-the-art machine learning approaches i.e. deep-learning. We will produce consistent, high resolution time series (starting in 2017) with detailed, ground-validated and spatially-refined time series of forest productivity. We will also explore the potential of recently proposed kernel-methods for deriving productivity indicators and compare these with classical metric approaches.