Hyperspectral in situ remote sensing of water surface nitrate in the Fitzroy River estuary, Queensland, Australia, using deep learning

TL;DR

This study tackles the challenge of estimating nitrate, a colorless nutrient, from optical measurements by leveraging time-series data of in-situ hyperspectral reflectance and nitrate concentrations at the Fitzroy River estuary. It treats the nitrate signal as indirectly linked to optically active water-quality parameters (e.g., TSS and CDOM) and uses a conditional denoising diffusion probabilistic model (cDDPM) to learn the joint distribution $P(NO_3^-, TSS, DOC \,|\, Rrs, S)$. The approach achieves a strong fit on held-out data with $R^2 = 0.86$ and $RMSE = 0.03\ \mathrm{mg/L}$, demonstrating the feasibility of predicting surface nitrate from spectral data and salinity. The method offers a pathway to operational nitrate mapping from spaceborne hyperspectral sensors, enabling up-scaling to satellites like PACE, EnMAP, and DESIS, thereby supporting nitrate management for the GBR lagoon.

Abstract

Nitrate ($\text{NO}_3^-$) is a form of dissolved inorganic nitrogen derived primarily from anthropogenic sources. The recent increase in river-discharged nitrate poses a major risk for coral bleaching in the Great Barrier Reef (GBR) lagoon. Although nitrate is an optically inactive (i.e., colourless) constituent, previous studies have demonstrated there is an indirect, non-causal relationship between water surface nitrate and water-leaving reflectance that is mediated through optically active water quality parameters such as total suspended solids and coloured dissolved organic matter. This work aims to advance our understanding of this relationship with an effort to measure time-series nitrate and simultaneous hyperspectral reflectance at the Fitzroy River estuary, Queensland, Australia. Time-series observations revealed periodic cycles in nitrate loads due to the tidal influence in the estuarine study site. The water surface nitrate loads were predicted from hyperspectral reflectance and water salinity measurements, with hyperspectral reflectance indicating the concentrations of optically active variables and salinity indicating the mixing of river water and seawater proportions. The accuracy assessment of model-predicted nitrate against in-situ measured nitrate values showed that the predicted nitrate values correlated well with the ground-truth data, with an $R^2$ score of 0.86, and an RMSE of 0.03 mg/L. This work demonstrates the feasibility of predicting water surface nitrate from hyperspectral reflectance and salinity measurements.

Figures (5)

• Figure 1: Location of the study site. Inset is a photo of the HydraSpectra instrument mounted on the pylon. The background image is displayed using a Sentinel-2 true-color composite, with Bands 4, 3, and 2 mapped to the red, green, and blue color channels, respectively.
• Figure 2: (a) The HydraSpectra instrument is equipped with a solar irradiance sensor, three skylight sensors, four water-leaving radiance sensors in the nadir and tilted directions, and two true-colour cameras facing the hemispherical sky and the water horizon. (b) True-colour images recorded by the sky camera. (c) True-colour images recorded by the water camera. (d) Randomly sampled example spectra of water-leaving reflectance recorded by the HydraSpectra instrument.
• Figure 3: The relationship between remote sensing reflectance, optically active water quality variables, and the optically inactive nitrate concentration (TSS: total suspended solids; CDOM: coloured dissolved organic matter).
• Figure 4: Time-series measurements of water surface nitrate concentration (in unit of mg/L) and salinity levels (in unit of PSU) at the study site on (a) April 7th, 2024, and (b) April 15th, 2024.
• Figure 5: Accuracy assessment of model-predicted nitrate against in-situ measured nitrate values.