Efficient Aggregate Land Cover Queries with Cloud-Optimized Raster Formats

Semantic queries of Earth observation (EO) imagery, such as “Find the images with less than 5% cloud cover in this area between two dates”, rely on aggregating some kind of scene classification data. Given the millions of pixels in a typical image, this can be resource-intensive. If knowledge is required of only the relative distribution of classes, though, do we need to process every pixel? The size and shape (morphology) of natural features, when observed by high-resolution optical sensors such as Sentinel-2, mean that simple downsampling can be used to drastically reduce the processing required for such queries, with only small losses in accuracy. Given an error tolerance of 1%, memory usage can be reduced by 625× and query runtime by 12×, for areas of interest of 60km × 60 km. Using cloud native technologies such as Cloud Optimized GeoTIFF (COG), this can also lead to significant reductions in network and disk usage.