EarthObsNet: A comprehensive Benchmark dataset for data-driven earth observation image synthesis

Recently, there are attempts to expand the current usage of satellite Earth surface observation images to forward-looking applications to support decision-making and fast response against future natural hazards. Specifically, deep learning techniques were employed to synthesize Earth surface images at the pixel level. Those studies found that precipitation and soil moisture play non-trivial roles in Earth surface condition prediction tasks. However, unlike many well-defined and well-studied topics, such as change detection, for which many benchmark datasets are openly available, there are limited public datasets for the abovementioned topic for fast prototyping and comparison. To close this gap, we introduced a comprehensive dataset containing SAR images, precipitation, soil moisture, land cover, Height Above Nearest Drainage (HAND), DEM, and slope data collected during the 2019 Central US Flooding events. Deep-learning-based SAR image synthesis and flood mapping with the synthesized images were presented as sample use cases of the dataset. • EarthObsNet - a dataset for pixel-level Earth surface prediction was introduced. • The dataset contains geospatial, meteorological, surface condition, and remote sensing data. • Benchmarks with three experiment configurations and model architectures were provided. • EarthObsNet can support other downstream tasks, such as flood inundation mapping.