Analysis-ready data (ARD) has attracted growing interest because it enables immediate downstream use of satellite imagery. Despite advances in preprocessing, many acquired images remain unsuitable for ARD generation due to quality-degrading factors. This study proposes a deep learning framework that evaluates the overall scene-level suitability of raw satellite images for ARD generation rather than assessing individual factors separately. The framework employs two independent CoAtNet models to classify (1) land cover types and (2) environmental conditions, both of which strongly influence ARD suitability. Their outputs are integrated into three suitability levels: high, moderate, and low. For rapid and computationally efficient assessment, the models are trained using KOMPSAT-3/3A browser images as practical proxies for raw imagery at the scene scale. Although browser images do not fully represent fine-grained radiometric and geometric properties, they provide reliable indicators for suitability evaluation. Experimental results demonstrate strong performance, achieving overall accuracies of 93.13 % for land cover classification and 95.41 % for environmental condition classification. The integrated suitability assessment reached an overall accuracy of 97.32 %, effectively identifying low-quality scenes inappropriate for ARD generation. The proposed framework offers an efficient, interpretable, and reliable pre-screening approach to support automated and cost-effective ARD workflows.



Keywords : Satellite scene classification, Analysis-ready data (ARD), CoAtNet, Land cover classification, Environmental condition classification