GhalamClouds: Remote sensing images for Kazakhstan and clouds segmentation using SAT-UNet


Aimyshev D. Tulegenov B. Smadyarov B. Nurzhakyp D.
1 January 2026Elsevier Ltd

Advances in Space Research
2026#77Issue 11 - 20 pp.

Cloud detection and segmentation represent a fundamental step in the analysis of optical remote sensing data. However, the performance of deep learning models heavily depends on the availability of task-specific datasets, which are often lacking. To address these challenges, we introduce two versions of the GhalamClouds dataset, derived from KazSTSat imagery, containing over 40,000 manually annotated cloud patches across six spectral channels and diverse landscapes of Kazakhstan. We also propose Sat-UNet, a U-Net backbone augmented with a spatial attention block designed to refine skip connections. Extensive experiments demonstrate that proposed model confidently competes with state-of-the-art models with significantly less parameters. Ablation studies further show that the proposed attention mechanism consistently outperforms the baseline UNet as well as SE and CBAM modules, with notable gains in F1 score, AUC, and Accuracy. We expect that GhalamClouds and Sat-UNet will provide valuable resources for advancing cloud segmentation research and improving model robustness in remote sensing applications.

Attention , Deep learning , Image processing , Remote sensing , UNet

Text of the article Перейти на текст статьи

Ghalam LLP, 89 Turan Avenue, Astana, 010000, Kazakhstan
Nazarbayev University, 53 Kabanbay Batyr, Astana, 010000, Kazakhstan

Ghalam LLP
Nazarbayev University

10 лет помогаем публиковать статьи Международный издатель

Книга Публикация научной статьи Волощук 2026 Book Publication of a scientific article 2026