With the explosive growth of satellite-based remote sensing missions, the image data generated on satellites has exponentially increased. However, limited space-to-ground communication bandwidth is a critical bottleneck, hindering swift transmission of large volumes of valuable image data to ground stations. Environmental noise during observations further degrades data quality, reducing its practical utility. Consequently, efficiently selecting and prioritizing satellite image transmission based on ground station queries has become a key challenge in space information retrieval. In this paper, we introduce a groundbreaking Space-Oriented Image Retrieval (SOIR for short) task, aimed at retrieving high-quality satellite images using ground-transmitted queries. To tackle this task, we categorize common satellite noise perturbations into two types: coarse and fine-grained perturbations. Based on this, we employ a sophisticated data augmentation strategy that constructs different types of perturbation masks to simulate varying levels of perturbation. After this, we introduce a Multi-task based Dual Generative Model (DGM for short) designed to rapidly filter high-quality satellite image data in response to image queries, significantly improving data retrieval and transmission. We conduct comparative experiments against a range of established image retrieval models. The experimental outcomes reveal that DGM offers substantial benefits in both retrieval efficiency and effectiveness.