GhalamClouds v2 - a dataset for semantic cloud segmentation based on KazSTSat imagery archive

GhalamClouds v2 is a collection of images from the KazSTSat earth observation satellite, manually annotated for semantic cloud segmentation. Images are selected from the period between 2023–2024 over Kazakhstan territory. The sensors operate at 686 km altitude with 22 m ground sample distance per pixel. GhalamClouds v1 was published previously on Figshare in 2025 (DOI: 10.6084/m9.figshare.29125640). 

This version is not divided into subsets. Original scenes of 14456×1800 pixels were padded with black margins to 14592×1920 — dimensions evenly divisible by 384 — and split into 384×384 patches. Four spectral channels are provided: Near-Infrared (nir), Red (red), Green (green), and Blue (blue), along with cloud mask ground truth (gt). Red Edge and Coastal Blue channels from v1 are excluded in this release.

The dataset contains a training split of 220 scenes (41,800 patches) and a test split of 77 scenes. Test images are distributed as patches for inference; ground truth masks are provided as full 14592×1920 scenes. Evaluation requires stitching predicted patches back into full scenes before comparison against ground truth, ensuring consistent benchmarking conditions across models.

Band alignment was corrected in this version. v1 alignment was performed per patch, which introduced misalignment artifacts; v2 alignment is performed per scene. Ground truth annotations are unchanged from v1.