Quantifying phytoplankton biomass and sediment in river plumes along the Agulhas Bank using remotely sensed data with deep learning techniques

Pillay, Humeshni

Quantifying phytoplankton biomass and sediment in river plumes along the Agulhas Bank using remotely sensed data with deep learning techniques

Thesis / Dissertation

2023

Abstract

River plumes play a major role in land-sea connectivity by providing essential nutrients, sediment, and organic matter to marine environments, and maintaining ecosystem function and habitat structure. Phytoplankton communities' reliance on riverine nutrient deposition to build up their biomass and suspended sediments released from terrestrial systems have been identified as major products of riverine outputs that impact pelagic and benthic ecosystems. Quantifying their concentrations in river plumes could improve our understanding of how anthropogenic activities, seasonality and climate change affect their concentrations in marine habitats and ecosystems over time. The need to monitor phytoplankton biomass and sediment loads in coastal systems has encouraged the use of remotely sensed data obtained from satellites. Remote sensing capabilities have advanced over the past few decades, along with increased computational efficiency, and sizeable open-access data pipelines. However, current satellite data products are not always appropriate for highly turbid, optically complex waters, along coastal regions where Atmospheric Correction (AC) is known to be challenging. The aim of this project was to use four rivers along the Agulhas Bank of South Africa to test satellite-derived Total Suspended Matter (TSM) and chlorophyll-a concentration ([Chl-a]) data products using a suite of available AC options against in situ [Chl-a] and TSM. The in situ data were collected through filtrations for fluorometric ([Chl-a]) and gravimetric (TSM) analyses. The performances of geophysical algorithms for TSM and [Chl-a] were assessed using remote sensing reflectance (Rrs) derived from standard Sentinel-3 Level-2 files, as well as two alternative approaches to AC, namely the Case-2 Regional CoastColour (C2RCC) processor, and the POLYnomial-based for MERIS (POLYMER). Thereafter, a regionally parameterised deep learning neural network (Multi-Layer Perceptron; MLP) model for retrieving TSM and [Chl-a] was developed and evaluated in the context of its application to Sentinel-3 satellite data. The MLP model is trained on a synthetic dataset of Rrs parameterised using the in situ ranges of TSM and [Chl-a]. The MLP model was evaluated using the three mentioned ACs as the model requires Rrs uncontaminated by the atmosphere. The regional MLP model was separated into a model containing the full range of TSM in situ values and a constrained version. POLYMER's [Chl-a] and C2RCC's TSM were the best geophysical products over the region when assessing the accuracy of retrievals to the match-up in situ dataset. However, AC is a major concern in regions where land-based fynbos biogenic burning is common. As a result, the current [Chl-a] satellite data product performances were not optimal, especially when mapped over the open ocean region. For the developed [Chl-a] MLP algorithm, the constrained and full MLP models were similar, with C2RCC's AC producing the best results. However, the constrained TSM MLP model's results vastly outperformed the full MLP model, with no clear delineation for the best-performing AC. The use of deep learning models shows promising results over the optically complex Agulhas Bank River plumes, however, the effectiveness of the MLP technique may be dependent on the variability of in situ data (i.e. phytoplankton size) used to create the training synthetic dataset, as well as the quality of Rrs applied to the MLP models. Thus, more in situ data is required to develop flexible, regional algorithms for highly turbid waters along South Africa's atmospherically variable coastlines.

Keywords

Biological Sciences

Reference:

Collections

Masters

Full item page