Abstract
The impact of coastal aquaculture on water quality should be assessed for operational monitoring and environmental management. Ammonium (NH 4 + ) is among the most vital nutrients originated from floating cage farming. Up till now, field sampling and laboratory analysis are routinely applied, a costly and time-consuming process. Satellite technology and machine learning techniques have been proved a useful tool for the operational monitoring of aquatic environment. Multispectral satellite images often show strong collinearity between the spectral bands, which should be reduced. The correlation coefficient, by itself, is not sufficient for the study of collinearity. Additional criteria must be established before one can proceed with machine learning techniques on satellite imagery. Based on in-situ seasonal measurements, the corresponding pixel values from the remotely sensed data were extracted to operate as the predictors. Stepwise regression analysis was applied to identify the most important variables for NH 4 + estimation, and variance inflation factor (VIF) was utilized as an indicator to assess collinearity among the predictors. As a result, spectral bands B2, B3, B4, B11, and max depth were selected as predictors. Different regression machine learning algorithms were tested, and based on the pairwise comparison method, gradient boosting (catboost) and support vector machine (SVM) were ranked higher for NH 4 + estimation.