Seeing through clouds. "Over the last few years, several cloud removal techniques have been developed: Sentinel-2 Cloudless, Mapbox Cloudless Atlas and Google's Cloudless Satellite Map. These techniques are sifting through multiple years of images to create cloud-free scenes. This works fine if the goal is to improve the aesthetics of the satellite images, but the trade-off is recency. In applications like ours, where we rely on the most recent images to monitor changes on the land, these techniques cannot be used."
"Clouds can be broadly classified into two types: dense, and thin or cirrus clouds. Dense clouds do not allow the penetration of visible spectral radiation from the ground and tend to cast a shadow on the ground."
"Thin or cirrus clouds on the other hand are transparent or semi-transparent clouds. Most spectral bands can partially see through these clouds. This is key to accurate de-clouding with our model."
"Our main contribution is to create a training dataset based on (cloud, no-cloud) pairs for the same geography but from different days. The machine learning model then learns to re-construct the no-cloud sample from the cloud sample."
"The basic assumption is that between cloud and no-cloud scenes only cloud pixels will change. This simple approach worked very well for us."
Seeing Through Clouds#
solidstatelife #
ai #
computervision #
satelliteimagery #
agriculture #
clouds 
Our novel machine learning approach for seeing through thin and cirrus clouds
eng.ruumi.io