Aircraft contrails are an important source of aviation's non-CO2 climate effects, and fine-scale satellite observations provide a means of characterizing their spatial structure. Distinguishing narrow, adjacent, and overlapping contrails from complex backgrounds requires reference data with explicit pixel-level annotations. We present ContrailStruct30, a three-band thermal-infrared contrail segmentation dataset derived from Sustainable Development Science Satellite 1 (SDGSAT-1) observations at 30 m nominal sampling. The dataset contains 19,455 patches of 256 × 256 pixels from 819 source-scene entries acquired on 305 dates between August 2022 and September 2024, including 9,023 positive patches and 10,432 deliberately selected difficult negatives. Each sample provides an enhanced RGB image, corresponding three-band source values stored as uint16, and a human-corrected binary mask. Construction began with 300 manually annotated patches identified through ADS-B-assisted searches and continued through four rounds of model-assisted annotation, manual correction, and expert review. Fixed training, validation, and test partitions separate both source scenes and acquisition dates. As an application baseline, we developed a MaxViT–U-Net segmentation model and compared it with three public models under a shared training protocol to evaluate foreground recovery, centerline agreement, and difficult-background discrimination. Across three training seeds, the baseline achieved a mean positive-patch test Dice of 0.8230 ± 0.0056 (mean ± standard deviation), with strong foreground overlap and centerline agreement in the common evaluation. By combining fine-scale thermal-infrared observations, diverse contrail morphologies, expert-reviewed labels, and corresponding image representations, ContrailStruct30 provides a reusable resource for extracting fine contrail structures and developing and evaluating high-resolution segmentation methods. The dataset presented in this study corresponds to Version 3, which is publicly available via Figshare at <a href="https://doi.org/10.6084/m9.figshare.33435496.v3" target="_blank" rel="noopener">https://doi.org/10.6084/m9.figshare.33435496.v3</a> (Yu et al., 2026).