The spatial resolution, wide swath and revisit cycle, and open data policy of Sentinel-2 imagery provides the opportunity to monitor landscape changes effectively. However, the presence of cloud and their shadows inhibit monitoring efforts. Often useable data exists within a scene; the inclusion of these cloud-free portions increases the amount of data available. Within Indufor’s monitoring routines, localised cloud masks are routinely applied, to allow cloudy and cloud-free imagery to be processed.
Several masking routines exist, we introduce three, with all producing spatial masks of the cloud. For this example, we apply Indufor’s cloud-shadow algorithm so that the cloud and their associated shadow are masked. This improvement allows cloudy scenes to be analysed, or parts of each to be combined to create cloud-free mosaics. The first of this set shows Indufor’s cloud and shadow mask, produced by combining spectral bands to identify cloud. The shadows are masked by using the solar zenith (height) and azimuth (direction) angles available in the Sentinel-2 metadata to project a likely shadowed area from the detected cloudy pixels. Indufor’s mask is optimised to detect clouds in tropical countries where cloud types and heights vary across the landscape.
The second mask CDI developed by [Frantz et al. (2018)] uses a parallax approach – where clouds are identified using their position relative to the sensor. The final method S2 Cloudless was developed by ESA and uses machine learning to detect cloud.
The following interactive demo compares all cloud masks around Bartica, Guyana. The Sentinel 2 scene selected is the ‘best’ cloud-free scene available in the Sentinel 2 catalogue!
The cloud assemblies are typical of this location and include ‘popcorn’ cloud and cirrus over the rainforest.
The performance of the masks over the Sentinel scene vary, Indufor’s mask is less effective at detecting cirrus type cloud, but does detect popcorn cloud well. The ESA S2 Cloudless machine learning approach performs better in this example, capturing both the popcorn cloud and cirrus well. Where both of these approaches are problematic is the false-positive labelling on bright land features, namely deforested mining areas, as cloud. This situation drove the development of the CDI algorithm, which successfully separates the ground features from cloud, but does not capture cloud, particularly cirrus, as well as the other approaches.
All cloud masking approaches may be improved by applying Indufor’s shadow detection algorithm as a secondary step post detection. When combined this solution links well with indufor’s automated area review processes by including all useable imagery.