|Paper title||Addressing the uncertainties associated with remotely sensed chlorophyll-a in oligotrophic and mesotrophic lakes and reservoirs|
|Form of presentation||Poster|
Over the last two decades, the primary focus of the development and application of remote sensing algorithms for lake systems was the monitoring and mitigation of eutrophication and the quantification of harmful algae blooms. Oligotrophic and mesotrophic lakes and reservoirs have consequently received far less attention. Yet, these systems constitute 50 – 60% of the global lake and reservoir area, are essential freshwater resources and represent hotspots of biodiversity and endemism.
Uncertainties associated with remote sensing estimates of chlorophyll-𝘢 (chla) concentration in oligotrophic and mesotrophic lakes and reservoirs are typically much higher than in productive inland waters. Uncertainty characterisation of a large 𝘪𝘯 𝘴𝘪𝘵𝘶 dataset (53 lakes and reservoirs: 346 observations; chla < 10 mg/l, dataset median 2.5 mg/l) shows that 17 algorithms, either recently developed or already well established, have substantial shortcomings in retrieval accuracy with logarithmic median absolute percentage differences (MAPD) > 37% and logarithmic mean absolute differences (MAD) > 0.60 mg/l. In the case of most semi-analytical algorithms the chla retrieval uncertainty was mainly determined by phytoplankton absorption and composition. Machine learning chla algorithms showed relatively high sensitivity to light absorption by coloured dissolved organic matter (CDOM) and non-algal pigment particulate absorption (NAP). In contrast, the uncertainties of red/near-infrared (NIR) algorithms, which aim for lower uncertainty in the presence of CDOM and NAP, were linked to the total absorption of phytoplankton at 673 nm and variables related to backscatter. Red/NIR algorithms proved to be insensitive to chla concentrations below 5 mg/l.
Bayesian Neural Networks (BNNs) for OLCI and the Sentinel-2 Multispectral Instrument (MSI) were developed as an alternative approach to specifically address the uncertainties associated with chla concentration retrieval in oligotrophic and mesotrophic inland water conditions (data from > 180 systems, n > 1500). The probabilistic nature of the BNNs enables to learn the uncertainty associated with a chla estimate. The accuracy of the provided uncertainty interval can be consistently improved when as little as 10% of the training data are set aside as a hold-out set. The BNNs improve the chla retrieval when compared versus established and frequently used algorithms in terms of performance over the expected training distribution, when applied to independent regions outside of those included the training set and in the assessment with OLCI and MSI match-ups.