An Open-Access Cross-Modal Forest Benchmark Training Dataset With Sentinel-1 and Lidar data

Forest height estimation is critical for applications such as forest health monitoring, carbon sequestration, timber supply, biodiversity assessment, and ecosystem dynamics. The growing interest in machine learning, especially deep learning, has the potential to significantly enhance the automatic interpretation of satellite images for forest height estimation. Sentinel-1, with its regular global coverage of Synthetic Aperture Radar (SAR) data, coupled with deep learning techniques, presents a valuable opportunity to create automated tools for forest monitoring. The success of such AI-driven methods relies heavily on the availability of high-quality public benchmark datasets that include large volumes of annotated training and test data. To address the current gap in forest SAR datasets, this paper introduces a new benchmark dataset, comprising pre-processed Sentinel-1 polarimetric interferometric data annotated with forest height information derived from space-based LiDAR data. This benchmark database includes a wide range of SAR data acquired over various interferometric and temporal baselines, covering diverse forest types and regions globally. The SAR data are provided as geocoded 4×4 polarimetric-interferometric covariance matrices, which highlight key features sensitive to forest structure and height, thus facilitating the development of machine learning models for forest monitoring.