Ground Based SAR Data Obtained With Different Combinations of Bandwidth and Step Size

GBSAR parameters

The dataset consists of 10935 raw Ground Based SAR (GBSAR) measurements obtained with GBSAR-Pi [1]. The measurements were conducted using developed GBSAR-Pi with different combinations of parameters. The combinations included five different step sizes (0.5, 1, 1.5, 3, and 3.5 cm), and three bandwidth cases (300, 600, and 1300 MHz).

GBSAR-Pi works in stop-and-go mode. GBSAR sensor is FMCW (Frequency Modulated Continuous Wave) radar with central frequency at 24 GHz. Total aperture remained the same throughout the measurement process (50 cm). FMCW radar periodically transmits sawtooth signal every 155 ms. Wave polarization of FMCW radar is set to vertical.

Classes

The observed scenes included three ceramic cups with 8 cm diameter placed 0.5 m away from the radar while the distance between them was 6 cm. The cups were either empty or contained aluminium or rubber object inside. The position of the cups remained the same throughout all measurements.
Hence, the scenes could not be distinguished by cups' visual properties nor change of their positions but only based on the different reflectance of the target object set inside of the cup which is why deep learning classification had to be trained on radar data rather than optical images.

In the dataset setting a single class represented a scene in which there is either nothing in the ceramic cups or an aluminium or rubber object set inside one of the three cups. Scenes in which multiple objects were set in multiple cups are not recorded. Hence, there is total of 7 classes:

ker0 - three empty cups,
ker1 - aluminium object in first cup (shown in Figure \ref{different_bandwidth}),
ker2 - aluminium object in second cup,
ker3 - aluminium object in third cup,
ker1g - rubber object in first cup,
ker2g - rubber object in second cup,
ker3g - rubber object in third cup.

Data format

GBSAR matrices are given in .npy format. The dimensions of the matrices depend on the number of steps used in that measurement. Hence, since the number of frequency points did not change throughout the measurements, size of one axis remains 1024 in all data points, while the size of other axis represents the number of steps.

Naming
{class_name}_{bandwidth}_{step size}_{number of steps}_{id}.npy

This work was supported in part by Croatian Science Foundation (HRZZ) under the project number IP-2019-04-1064.

[1] Kačan M, Turčinović F, Bojanjac D, Bosiljevac M. Deep Learning Approach for Object Classification on Raw and Reconstructed GBSAR Data. Remote Sensing. 2022; 14(22):5673. https://doi.org/10.3390/rs14225673