Multi-Scale Image Dataset for Semantic Segmentation of Geotechnical Anomalies in Mining Infrastructure

## Overview This dataset was developed to support research on the automated visualinspection and semantic segmentation of surface anomalies in mininggeotechnical structures. It combines images acquired at two complementaryobservation scales: 1. **Orthophoto dataset:** image tiles extracted from three aerial-photogrammetric surveys of an active mining operation, including waste-rock stockpiles, a tailings dam, and an open pit. 2. **Proximal-image dataset:** 1,529 images acquired during geotechnical inspections using smartphones and unmanned aerial vehicles (UAVs). The combination of aerial and proximal imagery enables the investigation ofboth area-wide monitoring and detailed inspection of geotechnical anomalies. ## Annotation classes The images were manually annotated for four types of visually identifiablesurface anomalies: - water accumulation or ponding;- erosion;- cracks;- surface ruptures. The annotations were prepared using the Roboflow platform and exported usingthe COCO JSON annotation format. The deposited files include the availableimages, annotation data, and supporting documentation describing the datasetorganization. ## Intended use The dataset is intended for non-commercial academic research involving: - semantic and instance segmentation;- computer vision for geotechnical inspection;- automated anomaly detection;- UAV-based monitoring of mining infrastructure;- human-in-the-loop inspection systems;- geotechnical risk assessment and decision support. The dataset may also be used to evaluate transfer-learning architectures andmulti-scale image-processing workflows for geotechnical applications. ## Data provenance and limitations The images originate from real geotechnical inspections andaerial-photogrammetric surveys conducted in an operational mining environment.The dataset consequently represents heterogeneous illumination, surface,perspective, scale, and acquisition conditions. The class distribution is imbalanced because water accumulation, erosion,cracks, and surface ruptures occur at different frequencies in operationalsettings. In particular, some anomaly classes contain substantially fewerexamples than others. Users should consider this imbalance when training,validating, and interpreting machine-learning models. The spatial resolution of the orthophotos is suitable for area-wideidentification of larger anomalies, but may be insufficient for detectingcentimeter-scale cracks. Proximal smartphone and UAV images provide greaterdetail for fine-scale anomaly characterization. This dataset should not be used as the sole basis for assessing the stabilityor safety of a geotechnical structure. Model predictions require validation byqualified geotechnical professionals. ## Access conditions The record metadata are publicly available. Access to the image and annotationfiles is restricted because the data were acquired at an active miningoperation and are subject to confidentiality and site-security requirements. Access may be granted for non-commercial academic research upon submission of: - the applicant's name and institutional affiliation;- a description of the proposed research;- the intended use of the dataset; and- an agreement not to identify, geolocate, redistribute, or use the images for commercial purposes. Access requests are evaluated by the dataset owners and may be declined whenthe proposed use is incompatible with the applicable confidentialityrequirements. ## Related research The dataset supports the research manuscript: *A Novel CNN Framework for Automated Mapping of Geotechnical Pathologies inMining Applications.* The article DOI will be added to this record after publication. ## Funding The development of this dataset was supported by the Goiás State ResearchSupport Foundation (FAPEG), grant number 202410267000774.