One of the primary uses of satellite imagery for many applications is to monitor change across landscapes. Randomly scattered cloud across images is one an obvious nuisance that frustrates this process, and often analysts must resort to using manual interpretation and digitising to record changes. This issue can be overcome using pixel-level processing techniques that mine individual images to extract the cloud-free parts, record the date of the pixel to produce a cloud-free composite within the date range specified. Narrowing of this date range has been made more feasible with a five-day revisit rate of Sentinel 2 satellites. This solution enables wall-to-wall mapping of cloud-prone regions to identify change, and has been adopted by Guyana Forestry Commission in their efforts to assist in mapping and reporting deforestation.
Using Guyana, a region known for its constant cloud cover, we demonstrate the effectiveness of the technique to create cloud-free composites and to detect landscape change. We use a stepwise approach that first screens and eliminates pixels affected by cloud and cloud shadow, then extracts either pixels from the clearest observation over forested areas or the median of the cloud-masked result elsewhere. Where pixels from a single observation were used, a date may be baked into the image (via an additional band), a useful addition that aids in determining the date of the change. This process was used to produce two image composites: the first from 2018 Q4 and the second from the 2019 Q4, visible in the interactive tool below. The final composites represent the clearest available pixels from each period.
The method demonstrates the potential of per scene pixel-level routines for the generation of cloud-free imagery. These composites are easily integrated into desktop GIS software and used to track landscape changes over time, and link well with Indufor’s other detection routines.
