|Paper title||ForestAI: Towards the next generation of forest monitoring services with Artificial Intelligence|
|Form of presentation||Poster|
Forests are a vital foundation for biodiversity, climate, and the environment worldwide. As existential habitats for numerous plant species and animals, forests are a driving factor for clean air, water, and soil. While an accelerated climate change and its impacts, such as extreme weather events, threaten forests in these functions, continuous monitoring of forest areas becomes more and more important. The relevance of managing forests sustainably is also emphasized in the Agenda 2030 of the Sustainable Development Goals, in which forests are directly linked to multiple SGD goals such as “Life on Land” or “Climate Action”. At present, however, maps of forests are often not up-to-date and detailed information about forests is often not available.
In this work, we demonstrate how Artificial Intelligence (AI), particularly methods from Deep Learning, can be used to facilitate the next generation of Earth Observation (EO)-services for forest monitoring. Relying on EO imagery from the Sentinel-2 satellites, we first discuss the importance of incorporating the multi-spectral and multi-temporal properties of this data source into Machine Learning models. Focusing on the challenge of segmenting forest types from EO imagery, we adapt and evaluate several state-of-the-art architectures from Deep Learning for this task. We investigate different architectures and network modules to integrate the high-cadence imagery (the constellation of the two Sentinel-2 satellites allows a revisit time of 5 days on average) into the Machine Learning model. In this context, we propose an approach based on Long-Short-Term-Memories that allows learning temporal relationships from multi-temporal observations. The comparison of our approach against mono-temporal approaches revealed a clear improvement in the evaluation metrics when integrating multi-temporal information.
We show how the proposed Deep Learning models can be used to obtain a more continuous forest mapping and thus provide accurate insights into the current status of forests. This mapping can complete and supplement existing forest mappings (e.g., from the Copernicus Land Monitoring Service). To that end, we provide a Deep Learning-based segmentation map of forests on a Pan-European scale at 10-meter pixel resolution for the year 2020. This novel map is evaluated on high-quality datasets from national forest inventories and the in-situ annotations from the Land Use - Cover Area Frame Survey (LUCAS) dataset. We finally outline how our approaches allow additional near-real-time monitoring applications of large forest areas outside of Europe. This work is funded by the European Space Agency through the QueryPlanet 4000124792/18/I-BG grant.