Optimization of Water Body Color Classification with Convolutional Neural Network Through Forel-Ule Scale Class Reduction

This study presents a method to optimize water color classification based on the Forel-Ule scale using a Convolutional Neural Network (CNN). The original 21-class system presents challenges such as high computational complexity, overlapping spectral characteristics, and class imbalance. A class reduction approach is proposed to group similar spectral categories into three ecologically meaningful classes: oligotrophic (clear blue), mesotrophic (greenish), and eutrophic (brownish). Using a dataset of 3,018 labeled water body images from EyeOnWater and implementing a CNN architecture trained on both the original and the reduced class schemes, the experimental results show that the reduced 3-class model achieved significantly higher accuracy (94.0%) compared to the original 21-class model (44.3%). These findings demonstrate that class reduction improves classification robustness, simplifies interpretation, and enhances practicality for real-world environmental monitoring.