Ensemble machine-learning-based framework for estimating total nitrogen concentration in water using drone-borne hyperspectral imagery of emergent plants: A case study in an arid oasis, NW China

In arid and semi-arid regions, water-quality problems are crucial to local social demand and human well-being. However, the conventional remote sensing-based direct detection of water quality parameters, especially using spectral reflectance of water, must satisfy certain preconditions (e.g., flat water surface and ideal radiation geometry). In this study, we hypothesized that drone-borne hyperspectral imagery of emergent plants could be better applied to retrieval total nitrogen (TN) concentration in water regardless of preconditions possibly due to the spectral responses of emergent plants on nitrogen removal and water purification. To test this hypothesis, a total of 200 groups of bootstrap samples were used to examine the relationship between the extracted TN concentrations from the drone-borne hyperspectral imagery of emergent plants and the experimentally measured TN concentrations in Ebinur Lake Oasis using four machine learning (ML) models (Partial Least Squares (PLS), Random Forest (RF), Extreme Learning Machine (ELM), and Gaussian Process (GP)). Through the introduction of the fractional order derivative (FOD), we build a decision-level fusion (DLF) model to minimize the regression results’ biases of individual ML models. For individual ML model, GP performed the best. Still, the amount of uncertainty in individual ML models renders their performance to be subpar. The introduction of the DLF model greatly minimizes the regression results’ biases. The DLF model allows to reduce potential uncertainties without sacrificing accuracy. In conclusion, the spectral response caused by nitrogen removal and water purification on emergent plants could be used to retrieve TN concentration in water with a DLF model framework. Our study offers a new perspective and a basic scientific support for water quality monitoring in arid regions.