Monitoring primary productivity in the Arctic Ocean waters has received increased interest in recent years due to the rapid environmental changes. Retreating and thinning sea ice, with increased light penetration, and favorable nutrient conditions, have opened new habitats for phytoplankton. Similarly to terrestrial plants, phytoplankton also uses photosynthesis in order to live and grow. Photosynthetic processes take place in the Chlorophyll-a (Chl-a) molecule. Therefore, the monitoring of Chl-a content may be an efficient way of retrieving information about the health status of the Arctic waters.
The absorption spectrum of the Chl-a is within the visible range of the electromagnetic spectrum, and hence it is retrieved with optical imaging sensors. There are several operational optical sensors, which are acquiring data over the Arctic, such as NASA’s operational Moderate Resolution Imaging Spectroradiometer onboard AQUA (MODIS- AQUA), Sentinel-2A and Sentinel-3.
In order to retrieve information about the Chl-a content from optical imaging satellite data, accurate, fast and robust algorithms are required. In the case of MODIS-AQUA, the state-of-art algorithm is the so-called OC3 model. However, it has been found that this algorithm often results in under- and overestimates of the Chl-a concentrations when applied at Arctic latitudes. In order to improve Chl-a content estimates, we propose to use alternative regression models, and in this paper, we will examine two new powerful regression models, namely, the Gaussian Process Regression model (GPR) and the Partial Least Squares Regression (PLSR) model.
GPR is known to be a flexible machine learning algorithm, with several advantageous properties. It has an analytical solution, a strong regression capability and excellent performance. In addition to the estimated Chl-a content, the output of the model also includes the predictive variance. Thus, the certainty level of the estimated Chl-a content can also be accessed in an automatic way.
The PLSR model has also been successfully applied for Chl-a content estimation for the optically challenging oceanic waters. PLSR is an iterative statistical model, which has several advantageous properties. It can cope with co-linearity between spectral features, and reduce the effect of noise in the data. It also can do regression with multidimensional outputs.
In this paper we demonstrate how maps of estimated Chl-a content in Arctic waters can be generated from remote sensing optical sensors using the GPR and PLSR models. In particular, we compare performances of these two algorithms to the state-of-the-art OC3 model for various sensors and spatial resolutions.