A scalable biodiversity observation and prediction platfom based on novel community sensors and artificial intelligence (Teilprojekt)

Biodiversity loss is one of the most urgent environmental challenges of the 21st century, yet our ability to observe, quantify, and predict biodiversity change remains limited. Conventional biodiversity monitoring is spatially sparse, labor-intensive, and difficult to scale, leading to substantial information gaps as climate and land-use change accelerate. Earth observation offers a unique foundation to address this challenge by delivering spatially explicit and temporally continuous information on environmental conditions across large regions.

This project develops a next-generation biodiversity observation and prediction platform where Earth observation data form the methodological backbone and are tightly coupled with advanced artificial intelligence. Multi-sensor satellite time series from Sentinel-1, Sentinel-2, Landsat 8/9, EnMAP, and Planet are processed into harmonized, analysis-ready data cubes that capture ecosystem structure and dynamics across space and time. Dense, equidistant Earth observation time series are generated using advanced preprocessing, artificial intelligence gap filling, and temporal aggregation.

To translate these Earth observation predictors into biodiversity information, the project uses a multimodal, multi-output deep learning framework that integrates heterogeneous inputs such as satellite imagery, time series, and scalar environmental variables. The model architecture enables information sharing across species through latent embeddings and can handle incomplete input data and partial biodiversity observations from rapid assessment tools. Transfer learning efficiently encodes Earth observation data into latent representations, reducing training data requirements and computational costs. This integrated Earth observation and artificial intelligence framework enables scalable, robust, and transferable biodiversity monitoring and prediction.