Creating and leveraging SAR benchmark datasets to facilitate machine learning application

Rapid advancements in ML have strongly boosted SAR research. Yet, ML applications in this field are constrained by the limited availability of valid SAR benchmark datasets. To address this gap, this paper proposes a standard framework for creating those. We introduce three categories for extracting SAR feature maps from SAR images and propose corresponding radarcoding methods to transform reference data into radar coordinates. The attribute quality of the generated SAR benchmark dataset is described using a local attribute table. It details the individual attributes of each SAR feature, while a global attribute table outlines the shared attributes across all SAR features. The quantitative quality of the dataset is assessed using quality control metrics. To facilitate online sharing and publication, the created SAR benchmark datasets are associated with using SpatioTemporal Asset Catalogs (STAC). We applied our framework in land use and land cover (LULC) classification. We generated two benchmark datasets from three Sentinel-1A SAR images of Groningen, The Netherlands, one for ML application demonstration and one for quality comparison. To validate the applicability of our SAR benchmark dataset, we employed four well-established ML models: HRNet, Unet, PSPNet, and DeepLabv3＋. Additionally, we designed an artificial neural network (ANN) model to further improve the LULC classification performance. The ANN model achieved an average precision of 0.93, a recall of 0.94, and an F1-score of 0.93. In conclusion, our proposed framework is capable to produce high-quality benchmark datasets, thereby effectively supporting ML applications in the SAR domain.