Optical remote sensing of water quality parameters retrieval in the Barents Sea
This thesis addresses various aspects of monitoring water quality indicators (WQIs) using optical remote sensing technologies. The dynamic nature of aquatic systems necessitate frequent monitoring at high spatial resolution. Machine learning (ML)-based algorithms are becoming increasingly common for these applications. ML algorithms are required to be trained by a significant amount of training data, and their accuracy depends on the performance of the atmospheric correction (AC) algorithm being used for correcting atmospheric effects. AC over open oceanic waters generally performs reasonably well; however, limitations still exist over inland and coastal waters. AC becomes more challenging in the high north waters, such as the Barents Sea, due to the unique in-water optical properties at high latitudes, long ray pathways, as well as the scattering of light from neighboring sea ice into the sensors’ field of view adjacent to ice-infested waters.
To address these challenges, we evaluated the performances of state-of-the-art AC algorithms applied to the high-resolution satellite sensors Landsat-8 Operational Land Imager (OLI) and Sentinel-2 Multispectral Instrument (MSI), both for high-north (Paper II) and for global inland and coastal waters (Paper III). Using atmospherically corrected remote sensing reflectance (Rrs ) products, estimated after applying the top performing AC algorithm, we present a new bandpass adjustment (BA) method for spectral harmonization of Rrs products from OLI and MSI. This harmonization will enable an increased number of ocean color (OC) observations and, hence, a larger amount of training data. The BA model is based on neural networks (NNs), which perform a pixel-by-pixel transformation of MSI-derived Rrs to that of OLI equivalent for their common bands. In addition, to accurately retrieve concentrations of Chlorophyll-a (Chl-a) and Color Dissolved Organic Matter (CDOM) from remotely sensed data, we propose in the thesis (Paper 1) an NN-based WQI retrieval model dubbed Ocean Color Net (OCN).
Our results indicate that Rrs retrieved via the Acolite Dark Spectrum Fitting (DSF) method is in best agreement with in-situ Rrs observations in the Barents Sea compared to the other methods. The median absolute percentage difference (MAPD) in the blue-green bands ranges from 9% to 25%. In the case of inland and coastal waters (globally), we found that OC-SMART is the top performer, with MAPD <30% in the visible bands. It is also demonstrated that the proposed BA method achieves better spectral consistency between OLI and MSI-derived Rrs products for varying optical regimes than previously presented methods.
Additionally, to improve the analysis of remote sensing spectral data, we introduce a new spatial window-based match-up data set creation method which increases the training data set and allows for better tuning of regression models. Based on comparisons with in-water measured Chl-a profiles in the Barents Sea, our analysis indicates that the MSI-derived Rrs products are more sensitive to the depth-integrated Chl-a contents than near-surface Chl-a values (Paper I). In the case of inland and coastal waters, our study shows that using combined OLI and BA MSI-derived Rrs match-ups results in considerable improvement in the retrieval of WQIs (Paper III).
The obtained results for the datasets used in this thesis illustrates that the proposed OCN algorithm shows better performance in retrieving WQIs than other semi-empirical algorithms such as the band ratio-based algorithm, the ML-based Gaussian Process Regression (GPR), as well as the globally trained Case-2 Regional/Coast Colour (C2RCC) processing chain model C2RCC-networks, and OC-SMART.
The work in this thesis contributes to ongoing research in developing new methods for merging data products from multiple OC missions for increased coverage and the number of optical observations. The developed algorithms are validated in various environmental and aquatic conditions and have the potential to contribute to accurate and consistent retrievals of in-water constituents from high-resolution satellite sensors.
Paper I: Asim, M., Brekke, C., Mahmood, A., Eltoft, T. & Reigstad, M. (2021). Improving chlorophyll-a estimation from sentinel-2 (MSI) in the Barents Sea using machine learning. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 14, 5529-5549. Also available in Munin at https://hdl.handle.net/10037/21863.
Paper II: Asim, M., Matsuoka, A., Ellingsen, P.G., Brekke, C., Eltoft, T. & Blix, K. (2022). A new spectral harmonization algorithm for landsat-8 and sentinel-2 remote sensing reflectance products using machine learning: a case study for the Barents Sea (European Arctic). IEEE Transactions on Geoscience and Remote Sensing, 61, 4200819. Published version not available in Munin due to publisher’s restrictions. Published version available at https://doi.org/10.1109/TGRS.2022.3228393. Accepted manuscript version available in Munin at https://hdl.handle.net/10037/28578.
Paper III: Asim, M., Matsuoka, A., Hafeez, S., Eltoft, T. & Blix, K. Spectral harmonization of Landsat-8 and Sentinel-2 remote sensing reflectance products for mapping Chlorophyll-a in coastal, lakes and inland waters. (Submitted manuscript).
ForlagUiT Norges arktiske universitet
UiT The Arctic University of Norway
Følgende lisensfil er knyttet til denne innførselen: