Estimation of chlorophyll-a in uncrewed aircraft systems imagery using autonomous surface vessel data with machine learning algorithms and feature selection techniques

Chlorophyll-a (Chl-a) is a critical biological indicator of the eutrophic state of water bodies, emphasizing the importance of its detailed characterization and continuous monitoring. This study evaluated the performance of 10 widely used Machine Learning (ML) algorithms in deriving the spatiotemporal distribution of Chl-a from uncrewed aircraft systems (UAS) imagery. Field data for Chl-a algorithm development were collected simultaneously using an Autonomous Surface Vessel (ASV), a hand-held radiometer, and a UAS in the Western Mississippi Sound (WMS). To ensure the algorithms were developed using accurate remote sensing reflectance data, the spectral response function of the UAS was applied to the radiometer measurements. An initial dataset comprising of 85 variables was compiled, including individual spectral bands, band ratios, vegetation indices, and three-band indices. Two feature selection techniques—Sequential Backward Floating Selection and Exhaustive Feature Selection—were employed to identify the optimal subset of variables. These techniques reduced the original d-dimensional feature space to a k-dimensional space by iteratively evaluating all possible feature combinations and selecting those that achieved the highest R2 scores for each ML algorithm. Model performance was further validated using a separate dataset and assessed through metrics such as Root-Mean-Square Difference (RMSD), Mean Absolute Relative Difference (MARD), and Average Percentage Difference (APD). Among the 10 ML algorithms tested, the extreme gradient boosting (XGB) algorithm demonstrated superior performance, achieving the highest R2 score of 0.848. Using a combination of two band ratios, three vegetation indices, and three three-band indices, the XGB algorithm achieved RMSD, MARD, and APD values of 0.538 μg/L, 0.407 μg/L, and 9.83 %, respectively. The XGB algorithm was subsequently applied to a time series of UAS imagery to generate Chl-a concentration maps, which revealed the spatiotemporal distribution of Chl-a across the study area. The methodology developed in this study provides a robust framework for monitoring Chl-a in the coastal waters of the WMS using UAS imagery. Furthermore, the techniques and findings provide valuable insights for advancing water quality monitoring efforts in other regions.