Taking the Long View: Enhancing Learning On Multi-Temporal, High-Resolution, and Disparate Remote Sensing Data

The progress made in computer vision and satellite technology has opened up new possibilities for observing societies and infrastructure. By analyzing vast amounts of high-resolution multi-temporal satellite data, decision-makers can gain valuable insights into population shifts, economic trends, and infrastructure performance. Nevertheless, challenges in this kind of imagery, such as varying image quality, imbalances in data collection between urban and rural areas, high costs, and the absence of image metadata, can impede the efficacy of these methods. In this work, we develop strategies for enhancing the performance of learning methods for high-resolution multi-temporal satellite imagery. We develop custom augmentation methods and inference techniques for identifying disparate image resolutions across historical imagery. We apply our generic techniques to the problem of detecting structures in longitudinal imagery, exhibiting modest but consistent performance improvements over baseline techniques. We then develop a case study analyzing the relationship between the expansion of electricity access and the growth in human settlements over time. We discover that across 1000 communities in Kenya over a decade, settings that received electricity access grew 15% more slowly than settings that did not receive electricity access. This non-intuitive and statistically robust finding challenges conventional wisdom about infrastructure provision and rural-urban migration, with potentially broad implications for assessing the impacts of infrastructure investments on rural lives and livelihoods. All data processing and modeling scripts are available at https://github.com/santiagocorrea/DeepSatGSD.