GPR Datasets for Machine Learning and AI

Why real data matters for GPR AI

Models trained exclusively on synthetic radargrams learn clean hyperbolas over homogeneous backgrounds — then fail on field data full of surface clutter, antenna ringing, soil heterogeneity, and overlapping reflections. Real datasets expose your model to this complexity from the start. The most robust published approaches combine synthetic pre-training with fine-tuning on real data: GPRbase provides the real component of that recipe.

Machine learning tasks on GPR data

Common research problems include: hyperbola detection and fitting (object localization), rebar detection and cover estimation in concrete, buried utility detection and material classification, layer boundary picking for pavements and stratigraphy, void and defect detection, and full radargram semantic segmentation. Each GPRbase dataset documents its targets, letting you select data matching your task — rebar-rich concrete scans for detection models, multi-utility corridors for classification, layered pavements for boundary picking.

Data structure: from DZT to tensors

A radargram is naturally a 2D array: traces (columns) × time samples (rows), with 16 or 32-bit amplitude values. Python libraries like readgssi load DZT files directly into numpy arrays, ready for PyTorch or TensorFlow pipelines. B-scans can be treated as images for CNN-based approaches, as sequences for recurrent models, or as raw signals for 1D convolutional architectures. The DZT header provides the physical calibration (time range, samples per scan) needed to convert predictions back to depths.

Preprocessing and augmentation

Typical preprocessing includes time-zero correction, background removal (mean trace subtraction), gain compensation, and normalization. Because GPRbase data is completely raw, you control the full pipeline — essential for reproducible research and for studying preprocessing sensitivity. For augmentation, real data can be combined with synthetic radargrams, cropped into patches, flipped horizontally, and noise-injected. Multi-antenna variety across datasets naturally augments frequency diversity.

Generalization across sites and antennas

A model that works on one site with one antenna often fails elsewhere. GPRbase datasets span multiple applications (concrete, utilities, geotechnical, archaeology, roads, geosciences), antennas (from 100 MHz to 2.7 GHz), and site conditions. This variety enables meaningful cross-domain evaluation: train on some datasets, test on held-out sites, and measure true generalization rather than site memorization.

Which datasets for which task

Not every dataset suits the same work. For hyperbola detection and fitting, reinforced concrete acquisitions offer numerous, regular targets of known geometry — and a welded mesh overlap zone provides the hard case, where responses overlap to the point of defeating automatic detection.

For segmentation and interface tracking, pavement profiles and surveys in favourable ground show continuous, legible layers. For void detection, one dataset covers karst cavities confirmed on site: it is the only one in the catalogue carrying that guarantee, and it is worth more than a larger unvalidated set.

For handling false positives, one profile combines buried utilities with an airwave — two neighbouring signatures, one of which corresponds to no buried target at all. Finally, the 3D grids, documented in spacing and origin, suit volumetric approaches, and georeferenced surveys suit work coupling detection with coordinates.

What these datasets are not

Better said plainly than left to be discovered: these datasets are not evaluation benchmarks. They are not annotated, carry no masks or bounding boxes, and ground truth exists for only some of them. Publishing them as a benchmark would be misleading.

They are documented field acquisitions, useful for experimentation, pre-training, robustness testing and qualitative validation. Acquisition parameters vary from one campaign to the next, because these files come from real jobs rather than from a protocol designed for research — which is both their limitation and their point. Annotation and verified ground truth are on the roadmap.

Citing and publishing with GPRbase data

CC BY-NC-SA 4.0 permits research use including peer-reviewed publications, with attribution. Cite the dataset identifier (dsXXX) and the GPRbase URL in your data availability statement. If you build annotated versions (bounding boxes, masks, picks), the share-alike clause means publishing them under the same license — growing the commons for the whole GPR ML community.

Accessing metadata programmatically

The catalogue is exposed as JSON, freely and without a key: /api/datasets.json returns every dataset with its application, antenna, frequency and targets. A Python and JavaScript client, published under the MIT licence with no dependencies, filters the catalogue, exports it to CSV and generates citations. It is available on GitHub and has its own DOI.

API documentation  ·  /api/datasets.json  ·  GitHub  ·  DOI