Multi-annual sub-pixel land cover mapping with TESSERA latent embeddings
Abstract
Mapping land cover in highly heterogeneous landscapes is challenging, and classifications have inherent limitations where the spatial resolution of remotely sensed data exceeds the size of small objects. Sub-pixel land cover fraction maps based on medium-resolution optical remote sensing data like Landsat or Sentinel-2 overcome this limitation but are often not consistently available across multiple years because creating temporally robust and comparable multi-class fraction models can be challenging. Latent embeddings from geospatial foundation models are promising as they represent high-dimensional, large-scale general-purpose spectral, temporal, and structural features derived from Earth Observation data that are largely independent of available high-quality observations at specific points in time. We here assessed the usability of TESSERA latent embeddings for machine-learning regression-based land cover fraction models and compared the performance to more conventional spectral-temporal metrics and spline coefficients as input features. We also assessed the suitability of these inputs to train temporally transferred and generalized fraction models. We found that TESSERA latent embeddings are a well-suitable input to land cover fraction mapping, performing slightly better than spectral-temporal-metrics-based models in many cases, both being notably outperformed by spline coefficients (MAE across classes and years 11.58 vs. 11.64 vs. 9.45). Spline coefficients also performed best in all temporally generalized and many temporally transferred models. However, models trained with TESSERA latent embeddings demonstrate the highest consistency of transferred models, highlighting their capability to handle data gaps. We suggest that TESSERA latent embeddings are a valuable input to fraction mapping where data availability across years is highly variable, but spline coefficients generally perform best when sufficient high-quality observations are available.
Related
- Training-Free Monitoring of Kenya with Pre-computed Satellite Embeddings: Lake Expansion, Construction Dating and Crop Retrieval with TESSERA
- Annual canopy height mapping in Italy from Landsat and multi-source LiDAR for disturbance and recovery monitoring
- Experimental Results of a Gas Fired Adsorption Heat Pump and Simulative Prediction of Annual Performance in a Multi-Family House
- Annual mean methane emissions from individual oil-gas point sources worldwide inferred from multi-satellite plume-imaging observations
- A transparent constrained stochastic framework for urban expansion forecasting and future service-network screening: Cape Town demonstration using annual building evidence, empirical land constraints, spatial hindcasting, and multi-horizon ensembles
BibTeX
@misc{schug2026multiannual,
title = {{Multi-annual sub-pixel land cover mapping with TESSERA latent embeddings}},
author = {Franz Schug and David Klehr and Jari Mahler and David Frantz},
howpublished = {EarthArXiv},
year = {2026},
month = sep,
doi = {10.31223/x59j7c},
url = {https://eartharxiv.org/repository/view/14912/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Multi-annual sub-pixel land cover mapping with TESSERA latent embeddings
AU - Franz Schug
AU - David Klehr
AU - Jari Mahler
AU - David Frantz
PY - 2026
DA - 2026/09/11
PB - EarthArXiv
DO - 10.31223/x59j7c
UR - https://eartharxiv.org/repository/view/14912/
AB - Mapping land cover in highly heterogeneous landscapes is challenging, and classifications have inherent limitations where the spatial resolution of remotely sensed data exceeds the size of small objects. Sub-pixel land cover fraction maps based on medium-resolution optical remote sensing data like Landsat or Sentinel-2 overcome this limitation but are often not consistently available across multiple years because creating temporally robust and comparable multi-class fraction models can be challenging. Latent embeddings from geospatial foundation models are promising as they represent high-dimensional, large-scale general-purpose spectral, temporal, and structural features derived from Earth Observation data that are largely independent of available high-quality observations at specific points in time. We here assessed the usability of TESSERA latent embeddings for machine-learning regression-based land cover fraction models and compared the performance to more conventional spectral-temporal metrics and spline coefficients as input features. We also assessed the suitability of these inputs to train temporally transferred and generalized fraction models. We found that TESSERA latent embeddings are a well-suitable input to land cover fraction mapping, performing slightly better than spectral-temporal-metrics-based models in many cases, both being notably outperformed by spline coefficients (MAE across classes and years 11.58 vs. 11.64 vs. 9.45). Spline coefficients also performed best in all temporally generalized and many temporally transferred models. However, models trained with TESSERA latent embeddings demonstrate the highest consistency of transferred models, highlighting their capability to handle data gaps. We suggest that TESSERA latent embeddings are a valuable input to fraction mapping where data availability across years is highly variable, but spline coefficients generally perform best when sufficient high-quality observations are available.
ER -