Can we estimate vegetation indices and biophysical parameters for forests using SAR & ancillary data and machine learning? Researchers from the EO4Landscape research team (Daniel Paluba and Přemysl Štych), in collaboration with Eurpean Space Agency (Bertrand Le Saux from the Φ-lab and Francesco Sarti), tried to answer this question in a recently published paper in Big Earth Data journal by Taylor & Francis.
The EO4Landscape research team from the Department of Applied Geoinformatics and Cartography is pleased to announce the publication of our latest study: "Estimating Vegetation Indices and Biophysical Parameters for Central European Temperate Forests with Sentinel-1 SAR Data and Machine Learning," authored by Daniel Paluba, Bertrand Le Saux, Francesco Sarti, and Přemysl Štych. This research has been published in theBig Earth Data journal and is accessible online at doi.org/10.1080/20964471.2025.2459300.
In this study, we address the limitations of optical satellite data, such as cloud cover, by utilizing Synthetic Aperture Radar (SAR) data from Sentinel-1. Our research demonstrates that SAR data, when combined with ancillary information and machine learning techniques, can effectively estimate key forest parameters, including the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Leaf Area Index (LAI), and Fraction of Absorbed Photosynthetically Active Radiation (FAPAR). The study focuses on both healthy and disturbed temperate forests in Czechia and Central Europe during the year 2021.
We developed a multi-modal time-series dataset using Google Earth Engine (GEE), incorporating temporally and spatially aligned Sentinel-1 and Sentinel-2 data, along with digital elevation model (DEM)-based features, meteorological variables, and forest type classifications. Our findings indicate that the inclusion of DEM-based and meteorological data enhances the accuracy of forest parameter estimations. Among the machine learning models evaluated, Extreme Gradient Boosting (XGB) and Random Forest regressors demonstrated high accuracy (R² between 70% and 86%) and computational efficiency, outperforming Automatic Machine Learning (AutoML) approaches. Furthermore, SAR-based estimations across Central Europe yielded results comparable to those within Czechia, highlighting their potential for large-scale applications. A significant advantage of SAR-based vegetation metrics is their ability to detect abrupt forest changes with sub-weekly temporal accuracy, providing up to 240 measurements per year at a 20-meter resolution.
This research underscores the potential of integrating SAR data and machine learning for comprehensive and timely forest monitoring, offering a robust alternative to traditional optical methods.
The best-performing ML models for each VI accessible on Hugging Face.
Example code, data and instruction for using the models on GitHub.
Multi-modal time series were generated with the MMTS-GEE tool, developed in our previous paper, titled "Identification of optimal Sentinel-1 SAR polarimetric parameters for forest monitoring in Czechia", DOI: 10.14712/10.142/23361980.2024.18