Real-time turbidity classification using coarse K-nearest neighbors for adaptive underwater optical wireless communication systems

Underwater optical wireless communication (UOWC) is attractive for high-speed underwater data transmission, but variations in water turbidity can severely degrade communication performance. This study presents a supervised machine learning approach based on the Coarse K-Nearest Neighbors (K-NN) algorithm to identify channel turbidity states in real time. An analytical intensity modulation/direct detection (IM/DD) system with NRZ-OOK modulation at 520 nm was used to generate a balanced dataset of 3,802 samples over distances from 5 to 100 m. The classifier uses two features, the transmission distance and the quality factor (Q-factor), both normalized with an absolute Z-score. Performance was evaluated with stratified fivefold cross-validation. The Coarse K-NN configuration ( K = 100, Euclidean distance, uniform weighting) achieved an overall accuracy of 95.50 % and an area under the ROC curve (AUC) of 0.9550. A linear Support Vector Machine trained on the same features showed that the classes are not separated by a linear boundary in this normalized feature space; Coarse K-NN outperformed it by about 40 percentage points in overall accuracy. Using both distance and Q-factor resolves the ambiguity at short ranges, where similar Q-factor values may correspond to different turbidity conditions. The approach is simple and suits adaptive power control in underwater optical systems. The study uses an analytical channel model; experimental validation is still needed.