Mapping total suspended solids (TSS) and dissolved organic carbon (DOC) in complex coastal waters using deep learning enhanced remote sensing
Satellite remote sensing enables the mapping of large-scale spatiotemporal variations of Water Quality (WQ) parameters, including Total Suspended Solids (TSS) and Dissolved Organic Carbon (DOC). Ground-based bio-optical measurements are used to quantify and evaluate the relationship between the remo...
Saved in:
| Main Authors: | , , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-12-01
|
| Series: | Ecological Informatics |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S1574954125002857 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Satellite remote sensing enables the mapping of large-scale spatiotemporal variations of Water Quality (WQ) parameters, including Total Suspended Solids (TSS) and Dissolved Organic Carbon (DOC). Ground-based bio-optical measurements are used to quantify and evaluate the relationship between the remote sensing reflectance (Rrs) and the in-situ WQ parameters. However, the non-linear relationship is strongly influenced by optically complex coastal regions. We focus on quantifying the non-linear Rrs – WQ relationship using a dense Deep Learning (DL) model trained on an augmented spectral library dataset covering estuarine and off-shore regions along New South Wales coastline, Australia. The spectral library consists of WQ parameters and Inherent Optical Properties (IOP) of water samples, including absorption and backscattering coefficients that drive a radiative transfer model to derive resultant Rrs. We incorporate uncertainty into the WQ−IOP measurements and simulate an augmented spectral library for training a DL model for aquatic Remote Sensing (DL-RS) applications. The model error analysis revealed a) an inherent Mean Absolute Error (MAE) of 7 % for TSS and 22 % for DOC, b) a predictive MAE of 75 % (TSS) and 31 % (DOC) in temperate waters, and c) an in-situ optical closure predictive MAE of 18 % and a Mean Percentage Error of ±5 % upon comparing forward modelled Rrs outputs from DL-RS-derived IOP and WQ parameters with atmospherically corrected Landsat-8 imagery (2013−2022). The DL-RS model demonstrated improved performance and added computational efficiency over existing inversion models. The model thus provides valuable insights into the dynamics of WQ parameters in optically complex waters influenced by diverse land cover catchments. |
|---|---|
| ISSN: | 1574-9541 |