Processed RFI Segmentation Datasets for Cross-Telescope Generalization Benchmark

The NSRT-low_el domain corresponds to low-elevation observation scenarios, while the NSRT-pulsar_sf domain corresponds to pulsar-related observation scenarios.The processing workflow includes reading the original observation or public source data, constructing image or dynamic-spectrum tensors, generating non-overlapping 512 x 512 patches, median subtraction, channel-level standardization, and producing binary RFI masks using AOFlagger-derived reference masks where applicable.

In each HDF5 file, arr0 stores the processed image or dynamic-spectrum tensor, and arr1 stores the corresponding binary RFI mask or label. The files can be read using common scientific Python tools such as h5py, NumPy, and PyTorch.The dataset is not associated with a single geographic study region. Instead, it is organized by telescope or data-source domain, including HERA, LOFAR, and NSRT.

The temporal information depends on the original source data. The NSRT-low_el data mainly originate from low-elevation observations in January 2024, and the NSRT-pulsar_sf data mainly originate from pulsar-related observations in July 2024. The temporal information for the HERA and LOFAR domains follows the corresponding Zenodo source records.

The submitted files are processed machine-learning datasets and do not preserve the full structure of the original observation files.The initial submission contains 10 HDF5 files with a total size of approximately 103 GiB. The HERA domain includes hera_strong_train.h5 and hera_strong_val.h5. The LOFAR domain includes lofar_strong_train.h5 and lofar_strong_val.h5.

The NSRT-low_el domain includes low_el_train_v3b.h5, low_el_val_v3b.h5, and low_el_ood_test.h5. The NSRT-pulsar_sf domain includes pulsar_sf_train.h5, pulsar_sf_val.h5, and pulsar_sf_test.h5. These files correspond to the training, validation, and test splits used in the associated experiments and can be used together with the related code repository for reproducibility.The masks and processed tensors may be affected by the quality of the original observations, the AOFlagger reference masks, the preprocessing and standardization procedures, and patch-boundary effects.

The binary RFI masks are intended for model training and evaluation, but should not be interpreted as error-free manual pixel-level ground truth. The main purpose of this dataset is to provide a unified, reproducible benchmark for cross-telescope RFI segmentation experiments rather than to replace the original observation data.