Remote sensing of phytoplankton biomass in oligotrophic and mesotrophic lakes: addressing estimation uncertainty through machine learning

Abstract

Phytoplankton constitute the bottom of the aquatic food web, produce half of Earth’s oxygen and are part of the global carbon cycle. A measure of aquatic phytoplankton biomass therefore functions as a biological indicator of water status and quality. The abundance of phytoplankton in most lakes on Earth is low because they are weakly nourished (i.e., oligotrophic). It is practically infeasible to measure the millions of oligotrophic lakes on Earth through field sampling. Fortunately, phytoplankton universally contain the optically active pigment chlorophyll-a, which can be detected by optical sensors. Earth-orbiting satellite missions carry optical sensors that provide unparalleled high spatial coverage and temporal revisit frequency of lakes. However, when compared to waters with high nutrient loading (i.e., eutrophic), the remote sensing estimation of phytoplankton biomass in oligotrophic lakes is prone to high estimation uncertainties. Accurate retrieval of phytoplankton biomass is severely constrained by imperfect atmospheric correction, complicated inherent optical property (IOP) compositions, and limited model applicability. In order to address and reduce the current estimation uncertainties in phytoplankton remote sensing of low - moderate biomass lakes, machine learning is used in this thesis. In the first chapter the chlorophyll-a concentration (chla) estimation uncertainty from 13 chla algorithms is characterised. The uncertainty characterisation follows a two-step procedure: 1. estimation of chla from a representative dataset of field measurements and quantification of estimation uncertainty, 2. characterisation of chla estimation uncertainty. The results of this study show that estimation uncertainty across the dataset used in this chapter is high, whereby chla is both systematically under- and overestimated by the tested algorithms. Further, the characterisation reveals algorithm-specific causes of estimation uncertainty. The uncertainty sources for each of the tested algorithms are discussed and recommendations provided to improve the estimation capabilities. In the second chapter a novel machine learning algorithm for chla estimation is developed by combining Bayesian theory with Neural Networks (NNs). The resulting Bayesian Neural Networks (BNNs) are designed for the Ocean and Land Cover Instrument (OLCI) and MultiSpectral Imager (MSI) sensors aboard the Sentinel-3 and Sentinel-2 satellites, respectively. Unlike established chla algorithms, the BNNs provide a per-pixel uncertainty associated with estimated chla. Compared to reference chla algorithms, gains in chla estimation accuracy > 15% are achieved. Moreover, the quality of the provided BNN chla uncertainty is analysed. For most observations (> 75%) the BNN uncertainty estimate covers the reference in situ chla value, but the uncertainty calibration is not constantly accurate across several assessment strategies. The BNNs are applied to OLCI and MSI products to generate chla and uncertainty estimates in lakes from Africa, Canada, Europe and New Zealand. The BNN uncertainty estimate is furthermore used to deal with uncertainty introduced by prior atmospheric correction algorithms, adjacency affects and complex optical property compositions. The third chapter focuses on the estimation of lake biomass in terms of trophic status (TS). TS is conventionally estimated through chla. However, the remote sensing of chla, as shown in the two previous chapters, can be prone to high uncertainty. Therefore, in this chapter an algorithm for the direct classification of TS is designed. Instead of using a single algorithm for TS estimation, multiple individual algorithms are ensembled through stacking, whose estimates are evaluated by a higher-level meta-learner. The results of this ensemble scheme are compared to conventional switching of reference chla algorithms through optical water types (OWTs). The results show that estimation of TS is increased through direct classification rather than indirect estimation through chla. The designed meta-learning algorithm outperforms OWT switching of chla algorithms by 5-12%. Highest TS estimation accuracy is achieved for high biomass waters, whereas for low biomass waters extremely turbid waters produced high TS estimation uncertainty. Combining an ensemble of algorithms through a meta-learner represents a solution for the problem of algorithm selection across the large variation of global lake constituent concentrations and optical properties.