September 2023

Most Earth Observation satellites use optical sensors to image the ground: that is, cameras that collect radiation in the visible (and often near infrared) band of the electromagnetic spectrum to create an image or set of images at several wavelengths. All suffer from the fact that clouds are highly reflective at such wavelengths, and so detail below clouds cannot be seen. This has driven research into processing techniques that "fill-in" detail below clouds as a means of dealing with the problem. Here we share our capabilities in this regard and discuss the pros and cons of approaches to handling clouds in images.

Cloud is always raised as one of the main limitations with optical satellite imaging over other options for aerial image capture, such as drones or low-flying aircraft. The repeatability, reliability and cost-effectiveness of satellites, as well as the quality of their sensors, are usually sufficient reasons for the satellite option to win out, even when spatial resolution may have to be compromised. Nevertheless, users of optical satellite imagery can still find cloud cover an issue, whether simply frustrating or actually business limiting.

A number of strategies can be followed for dealing with the issue, each of which has benefits and drawbacks which must be considered when choosing how to handle cloudy capture days.

Image mosaics and cloud masking

Until recently, only two main strategies could be followed:

1. Observe the area of interest over a period of time so that the whole region is imaged over multiple cloudy days, or a single day provides a cloud-free image.

2. Automatically detect clouds and the shadows they cause in an image, and mask them to ensure that only valid areas of the image are processed.

The first approach is only applicable when time is not pressing, and the required imagery can be formed over several satellite updates or just one cloud-free. In some cases, this may only be a few days, but with some satellites there may be 3, 5, 12 or more days between updates and this must be considered when choosing to wait for a cloud-free day, or several days with few clouds. Using multiple images may also have cost implications, certainly when commercial satellite imagery is used, but even free satellite data incurs (albeit minimal) download costs.

Example of how two Sentinel-2 images can be mosaicked together to eliminate clouds

As shown above, if all images contain some cloud, then a mosaic technique must be relied upon, which requires processing skills as well as Option 2 (accurate cloud detection). Mosaicking requires either the ability to reliably detect clouds and shadows within an image or the use of cloud masks, illustrated below. This relies upon the detection algorithm being accurate enough not to falsely report bright or white areas to be clouds or misinterpreting dark areas of the ground as shadows . Generally, the quality of such masks is high, making them a good option for many applications. 

Example of a time series of Sentinel-2 images showing their corresponding scene classification (darker reds show thin and thick clouds)

Gap Filling

More recently, imputation methods – that is, algorithms that estimate missing values in datasets – have been applied to the problem of “filling in” the gaps left in images when clouds are removed. Whilst risky, as the estimated missing pixels can only be based on their previous state over time, these techniques have been honed and shown to provide low enough error rates in some cases to allow usable reconstructed images.

Leading methods for gap filling include deep learning, geo-statistics (e.g., Kriging method), and advanced non-linear interpolation techniques (e.g. DINEOF, Data Interpolating Empirical Orthogonal Functions). All require time series data, and in some cases, very large historical datasets in order to reach the very high performance required before being considered for use by D-CAT. Our rigorous assessment of the most promising techniques, followed by optimisation to minimise error rates, has resulted in a DINEOF-based method that is capable of image reconstruction, given a time series of preceding images, with an error of <0.3% for images with up to 60% cloud cover.

Example of Sentinel-2 image (left), reconstructed (right) using a statistical approximation from a time series of images

Our gap filling method is available as an alternative to our standard mosaicking or cloud masking approaches, but only where customers are aware of the limitations and risks of filling in missing data in a world where sudden, and sometimes catastrophic, changes can occur which will not be predicted by such techniques. For the viewing of imagery where typical change over time is expected and previous time series data is available, our gap filling option offers quality RGB imagery with established low error bounds, whilst customers requiring high radiometric accuracy across multispectral images are not advised to use this technique.

Mean NRMSE values for the gap filling technique applied to Sentinel-2 wavebands (5% cloud cover, left; 60% cloud cover, right)

For more details about our products and services, please get in touch.



If you want to discuss any of our MRV data products or capabilities, just...

Get in Touch