SpectralEarth-MM: Towards a foundation model for next-generation land cover monitoring across spectra

Foundation models are transforming remote sensing by enabling general-purpose solutions that work across different sensor types and tasks. While RGB, multispectral, and radar domains have seen rapid advances, hyperspectral imagery—with its ability to capture hundreds of spectral bands—remains a largely untapped resource.

SpectralEarth-MM bridges this gap by bringing together:

• EnMAP (hyperspectral)
• Sentinel-2 and Landsat-8 (multispectral)
• Sentinel-1 (radar)

While still under development, the resulting model is currently trained on a comprehensive, multi-sensor dataset with a focus on land monitoring downstream applications.

At the heart of SpectralEarth-MM is a vision transformer architecture built for flexibility and scale. Specialized layers are used to handle the unique characteristics of each sensor type. The model is trained with self-supervised techniques to learn unified representations across data sources without requiring exhaustive labelling. This architecture enables the model to handle a wide range of land monitoring tasks with improved performance and adaptability.

The capabilities of SpectralEarth-MM are being tested on several critical use cases:

• Forest monitoring: Classifying tree species with higher accuracy using the rich spectral information from hyperspectral data.
• Agriculture: Identifying crop types at scale, supporting precision farming and agri-environmental policies.
• General land cover: Performing both patch-level and pixel-level classification across landscapes for consistent, high-resolution mapping.

By unifying hyperspectral, multispectral, and radar data in a single benchmark and foundation model, SpectralEarth-MM opens the door to truly multi-sensor, scalable solutions for Copernicus Land Monitoring and beyond. Its ability to support diverse tasks, from forest species mapping to crop classification, makes it a key step towards more general, automated, and accurate Earth observation systems.

The SpectralEarth-MM dataset lays the groundwork for the next generation of land monitoring methods for EvoLand beyond current prototypes—flexible, cross-sensor, and ready to capture the complexity of our changing planet.

SpectralEarth-MM is an ongoing research effort, and our team is currently finalising the details for publication. Stay tuned for more updates as we move closer to sharing the full methodology and results.