Stacked Ensemble with Machine Learning Regressors on Optimal Features (SMOF) of hyperspectral sensor PRISMA for inland water turbidity prediction

Abstract

Leveraging hyperspectral data across various domains yields substantial benefits, yet managing many spectral bands and identifying the essential ones poses a formidable challenge. This study identifies the most relevant bands within a hyperspectral data cube for turbidity prediction in inland water. Nine machine learning regressors Cat Boost, Decision Trees, Extra Trees, Gradient Boost, Light Gradient Boost (LightGBM), Recursive Feature Elimination (RFE), Random Forest, Support Vector Regressor (SVR), and Xtreme Gradient Boost (XGBoost) have been used to compute the feature importance of the hyperspectral bands for predicting turbidity. Random Forest has outperformed the other models with a mean absolute percentage error (MAPE) of 1.61%, and the R² of the linear fit is 0.96. Band 77, with a central wavelength of 1067.61 nm, is the most dominating band regarding feature importance. We have also developed a novel framework for turbidity prediction: Stacked Ensemble with Machine Learning Regressors on Optimal Features (SMOF). It employs a stacking ensemble of the nine regressors mentioned above with Random Forest as both base and meta-model, leveraging feature selection outputs. With this framework, the MAPE (%) reached 1.21, while the R² stood at 0.95. The present study also presents a simple statistical algorithm to detect noisy bands in the Hyperspectral Precursor of the Application Mission (PRISMA) data cube. The approach assesses quadrat-wise intra-band spatial coherence using Renyi's entropy thresholding for noisy band segregation. Radiometric calibration error and absorption due to water vapour are the two primary sources of noise within the data cube. Moreover, this research implements the open-source Water Colour Simulator (WASI) to simulate inland water spectra with varied proportions of turbidity. Overall, the study presents an approach to identify noisy bands and integrates the potential wavelengths for turbidity prediction of inland waters.

Keywords: Ensemble stacking; Feature selection; Image fusion; Machine learning; Spectral noise; Water quality.