Spatially and Temporally Consistent Smallsat-Derived Basemaps for Analytic Applications

Smallsat constellations are increasingly relied upon for mapping and monitoring needs in remote sensing. Planet's constellation provides near-daily visual + NIR coverage at 3-5m cellsize allowing MODIS-like temporal monitoring at a higher spatial resolution than Sentinel or Landsat. The strengths of smallsat data come with several challenges for large-scale earth observation tasks, however. Smallsats often have a narrower swath width, increased sensor variability, and decreased positional accuracy compared to traditional remote sensing platforms. As a result, observing large regions requires combining multiple observations with differing characteristics and quality. Even with rigorous calibration, atmospheric correction, and validation of instruments it is challenging to combine large numbers of independent images into a consistent and cloud-free dataset.

We have developed an empirical approach to create spatially and temporally consistent datasets from millions of input scenes. Our approach works by 1) ranking scenes and finding the best available data for each area and time of interest, 2) applying radiometric and atmospheric corrections to each scene, 3) normalizing each scene based on an a-priori model of expected reflectance to remove additional variability, and 4) matching values at scene boundaries to reduce seamline effects. We divide work and storage spatially based on a global tile grid, allowing us to distribute computation horizontally and create global-scale RGB+NIR datasets at ~2-5 meter cellsize. The resulting analytic basemaps are spatially and temporally consistent at the expense of rigorous radiometric accuracy. They provide a ready-to-use dataset without the need for extensive pre-processing and per-scene corrections. Analytic basemaps are well suited to many common use cases involving supervised classification, such as land cover classification, deforestation monitoring, and urban growth monitoring.