Thesis
Neural networks applied to ocean colour remote sensing for environmental monitoring
- Creator
- Rights statement
- Awarding institution
- University of Strathclyde
- Date of award
- 2023
- Thesis identifier
- T16633
- Person Identifier (Local)
- 201979683
- Qualification Level
- Qualification Name
- Department, School or Faculty
- Abstract
- Autotrophic algal organisms that perform photosynthesis are the basis of the marine food web and absorb up to 30% of the anthropogenic emitted CO2 (Gruber et al., 2007). The algal concentration can be measured by retrieving their chlorophyll a concentration from remote sensing. The light signal measured from satellites needs to be corrected for various atmospheric and water surface effects. After atmospheric correction, the water leaving signal can be isolated and different algorithms exist to retrieve chlorophyll a. In open waters, blue-green ratios perform well (O’Reilly et al., 1998). In coastal waters, other water constituents (dissolved matter and sediments) make both the atmospheric correction process and the chlorophyll a estimates harder as they alter the light signal. Current chlorophyll algorithms therefore tend to perform poorly for turbid coastal waters. To develop a better algorithm, a northwest European shelf seas matchup dataset is built by collecting in situ chlorophyll a and MODIS Aqua sensor data. Different neural network algorithms are developed to make the best possible estimates of chlorophyll a, using either the bottom of atmosphere reflectance Rrs, commonly used by other algorithms, or top of atmosphere reflectances, when no atmospheric correction was applied to the light signal. It is found that the uncorrected top of atmosphere signal produces better and more reliable estimates over the entire dataset which contains mainly nearshore samples. Small neural network architectures containing 3 hidden layers of 15 neurons show good performances. The randomness involved in what a single neural network produces is tackled by using an ensemble approach of ten networks. The use of the whole light spectrum from 412 nm to 2130 nm produced the best estimates and should lead future dataset creations that currently do not include all available spectral information. The neural network algorithm developed here for chlorophyll a works well for turbid coastal waters where other algorithms either fail or mask out the data after applying quality control flags. The greatest impact is likely to be for nearshore waters where turbidity tends to be greatest. The approach developed here for NW European shelf seas has potential to be extended to global scale operation if a suitable training data set can be collected in future. The process is repeated for a modelled dataset and shows almost perfect estimates of the three different water constituents that alter the light signal. Temperature can also be estimated with better performance using the same approach which allows creation of a single temperature algorithm that works during both day and night, for any month and position on Earth. The benefit of using an enhanced chlorophyll a algorithm is later evaluated from the eutrophication assessment point of view. Remote sensing techniques produce several orders of magnitude more data than the current in situ approach to assess the eutrophication status and should help policymakers in producing coherent and improved assessment of the environment.
- Advisor / supervisor
- McKee, David
- Resource Type
- DOI
- Date Created
- 2022
Relazioni
Articoli
| Thumbnail | Titolo | Data caricata | Visibilità | Azioni |
|---|---|---|---|---|
|
|
PDF of thesis T16633 | 2023-06-21 | Pubblico | Scaricare |