This was the input image for STORM. It was a wide-swath image covering more than 210,000 sq.km, including the Egyptian stretch of the Nile River Valley on the left to the Red Sea on the right. Â The Red Sea facilitated the shortest shipping route between Asia and Europe, carrying more than 15% of global trade by value. STORM would have performed a series of processing steps on this image to generate insights.
Image Characteristics:
Satellite images often develop artefacts due to the harsh space environment. These artefacts appear as grainy particles, reducing the usefulness of the data. Uncorrected, they could lead to very inaccurate interpretations. A well-established solution was denoising, typically performed in ground-based labs after the image download. STORM's denoising algorithm actively found and removed the noise in the image, delivering a consistent and error-free output.
Due to the configuration of imaging satellites, executing an imaging operation may require the satellite to re-orient from pointing away from earth to the imaging target. To generate accurate insights all images other than the target will need to be filtered out. STORM utilizes a Deep Learning classifier model running on the satellite's Vision Processing Unit (VPU) to assess image usability. As shown in the video here, this AI identifies the image type and then makes a decision: continue to the next step or purge the image and terminate.
Atmospheric effects like clouds and haze hindered electro-optical sensors. These effects caused insights generated from the sensors to produce false values when encountering clouds in the area of interest. To address this challenge, STORM utilized a cloud detection model to actively detect and segment clouds within the denoised image from Step 1.
When working with coastal regions, it was preferred to segment waterbody and land regions so that the eventual application had the right context. This was run on the denoised image from Step 1, before any end-use models ran on the land or water pixels. The advantage of this step was that it reduced the computation resources needed and the number of pixels to be processed in subsequent steps. Unlike conventional water segmentation models, a spectral ratio approach was used to identify water regions in the absence of the Near Infrared (NIR) band.
P.S: With a spatial resolution of 204 m, the River Nile is hard to discriminate from its riparian corridor. The actual river spans less than 1 pixel, and hence not detected as a waterbody.
We can classify any landmass into multiple categories which helps us understand land-use patterns in the area. By repeating this classification over time, changes alongside anthropogenic phenomena can be tracked. STORM ran a pixel-level classifier on the land pixels. A spectral ratio approach that leverages visible bands was used in the absence of the Near Infrared (NIR) band.
Within a minute of loading STORM onto the satellite, it generated the Cloud Mask, Water Mask, and Vegetation Classification. Together, this combination was called an insight. To achieve this, STORM utilized the VPU and CPUs available on the satellite itself. When viewed holistically, this image informed us about the presence of clouds, water bodies, and the vegetation cover across the imaged region.