ocean engineering

Clouds, Marine Satellite Data, and Algorithms

Stock, A.; Subramaniam, A.; Van Dijken, G.L.; Wedding, L.M.; Arrigo, K.R.; Mills, M.M.; Cameron, M.A.; Micheli, F. Comparison of Cloud-Filling Algorithms for Marine Satellite Data. Remote Sens. 2020, 12, 3313. https://doi.org/10.3390/rs12203313

Our world’s marine ecosystems are exposed to many anthropogenic (originating in human activity) pressures from climate change, as well as from fishing, pollution, and habitat destruction. Therefore, monitoring how marine ecosystems change over time is more critical than ever.
Climatologies (the study of our climate) based on satellite and ground observations indicate that, on average, more than two-thirds of the world’s oceans are covered by clouds at any given time.
Since the launch of the Sea-viewing Wide Field-of-view Sensor (SeaWiFS) in 1997, the continuous availability of moderate-resolution, multi-spectral ocean color and thermal satellite imagery has expanded our ability to study and monitor marine phenomena at broad spatial scales.
Because electromagnetic radiation at visible, near- and thermal-infrared wavelengths is absorbed and scattered by clouds, essential marine satellite data products have significant gaps, limiting the ability to observe phenomena with high spatial and temporal variability.


Several research teams have developed algorithms that can reduce the problems caused by cloud cover in marine satellite data. The most widely used algorithm is DINEOF (Data-Interpolating Empirical Orthogonal Functions), which has been applied and tested in various studies, and allows the interpolation of single or several, correlated oceanographic variables.


The 3 Research Questions:

Therefore, Dr. Stock and team aimed to answer three related research questions.

(1) First, how accurately can the existing and various new algorithms predict Chl a concentrations “under the clouds”?  Note that Chlorophyll a concentrations are an indicator of phytoplankton abundance and biomass in coastal and estuarine waters. They can be an effective measure of trophic status (productivity of aquatic ecosystems) and are a commonly used measure of water quality.

(2) Second, which of these algorithms can be recommended when considering accuracy, computational cost, and the manual effort required to use them?

(3) Third, for researchers interested in developing new methods, what (if any) are the shared properties of the best-performing algorithms?


Remote Sensing | Free Full-Text | Comparison of Cloud Cover Detection Algorithms on Sentinel–2 Images of the Amazon Tropical ForestRemote Sensing | Free Full-Text | Comparison of Cloud Cover Detection Algorithms on Sentinel–2 Images of the Amazon Tropical Forest

Image source: Rem et al. 2019 [open access]


Study Areas and Data

Dr. Stock and team answered these questions based on analyses in four study areas of different sizes and a globally representative range of phytoplankton and cloud dynamics: The southern Beaufort Sea, the southern Chukchi Sea, part of the open equatorial Atlantic, and the Gulf of Mexico.

To test different gap-filling algorithms, they combined mostly cloud-free images with clouds “transplanted” from other images. They then used the algorithms to reconstruct the pixels under the clouds for each combination of cloud-free image and transplanted cloud mask, while all other images (cloud-free or not) were left unmodified for use by the reconstruction algorithms.

In contrast to previous research using a small number of clouds transplanted to a few example images for this purpose, they processed 750 to over 8000 combinations of mostly cloud-free images and transplanted clouds in each study area. Therefore, they were able to quantify the algorithms’ performance in a wide range of realistic situations.

Image source: Stock et al. 2020 [open access]


So, in summary…

This study compared ten gap-filling algorithms in four oceanographically heterogeneous study areas, spanning coastal to offshore waters and Arctic to tropical latitudes; their Arctic study areas can be optically complex even offshore.

Dr. Stock and team found that applying even simple cloud-filling algorithms before calculating spatial means of Chl a substantially reduced the means’ errors. Prior cloud-filling can thus be a straightforward way to improve regional time series derived from marine satellite data. Overall, they found that the best algorithm depends on the study area, the purpose of the gap-filling, and the period covered by the images.

What is Random Forest? | IBM

Image source: Lee et al. 2018 [open access]

Random forests including predictors that allow spatiotemporal interpolation reconstructed individual pixel values most accurately. This result suggests that the continued development of interpolating supervised learning methods for filling data gaps in marine satellite imagery is a promising direction for future research.

Interactive Online Maps Make Satellite Ocean Data Accessible - Eos

Image source: EOS AGU 2018 [open access]


Leave a Reply

Your email address will not be published.