An Adaptive Spatiotemporal Approach for Monitoring Change At Scale with Planet Imagery

An unprecedented opportunity to monitor Earth has emerged with commercial high resolution constellations like PlanetScope. The constellation produces near daily 5 m observations for the globe, promising the opportunity to monitor change with high spatial and temporal specificity. But modern "all-available data'' approaches have needed ever larger distributed computing investments just to keep pace with the increasing velocity of moderate resolution data production. Naive implementation of the same methods to the dramatically larger PlanetScope archive would require an equally dramatic investment in computation. But the vast majority of Earth is relatively stable the vast majority of the time. Likewise, most of Earth's surface state variability does not require 5 m imagery to adequately characterize. Knowing when and where more data is necessary to meet an objective would allow monitoring algorithms to scale by making the data much sparser in most places, most of the time. Here we explore spatiotemporal adaptive sampling schemes which balance exploration and exploitation to increase the spatial and temporal density of observations within and around areas deemed suspicious based on metrics such as Shannon's entropy and spatial autocorrelation. We applied the method to detect heavy construction events across several cities with diverse ecological and development contexts: Gangneung, Korea, Jacksonville, Florida, and Muharraq Island, UAE. The adaptive implementation achieved comparable change detection accuracy to benchmark algorithms that used all-available data, but required substantially less data to do it. Image archives grow larger each day, making it increasingly difficult for Earth monitoring methods to utilize all-available data. Approaches that minimize data, like the adaptive sampling scheme demonstrated here, will become increasingly relevant as we attempt to characterize and monitor our dynamic planet at ever larger scales.