Hyperspectral satellite data collections have been successfully leveraged in many domains such as meteorology, agriculture, forestry, and disaster management. There is also a collection of publicly available satellite observation networks. However, gaps in scanning frequencies and inadequate spatial resolutions limit the capabilities of geoscience applications. In this study, we target the temporal sparsity of high-resolution satellite images. In particular, we propose a novel methodology to estimate high-resolution images between scheduled scans. Our model SATnet, falls broadly within the class of Generative Adversarial Networks. SATnet allows us to generate accurate high-resolution, high-frequency satellite data at diverse spatial extents. SATnet achieves this by learning relations between a sequence of high-resolution/low-frequency satellite imageries (from Sentinel-2) and an ancillary satellite image that is high-frequency/low-resolution (from MODIS). Our benchmarks demonstrate that SATnet outperforms existing approaches such as ConvLSTMs, Dynamic Filter Network, and TrajGRU with a PSNR accuracy of 31.82. We trained and deployed SATnet over a distributed storage cluster to support the high-throughput generation of imputed satellite imagery via query evaluations. Our methodology preserves geospatial proximity and facilitates the dynamic construction of satellite imagery at a particular timestamp for arbitrary spatial scopes.