Classification of modulation schemes used in underwater communication using artificial intelligence techniques

Underwater wireless optical communication systems are a communication technology that aims to achieve high-speed data transmission by using the visible light band of the electromagnetic spectrum, which is least affected by absorption, scattering, and turbulence caused by the inherent properties of water. Despite the relative advantages of the visible band, these effects still lead to signal attenuation, thereby limiting system performance and reducing communication reliability. Automatic modulation classification is a technique that aims to determine the modulation type of a received signal at the receiver without requiring prior information. In recent years, various deep learning–based approaches have been developed for modulation classification in underwater wireless optical communication, achieving successful results particularly at high signal-to-noise ratio levels. However, these methods are known to show limited performance under challenging conditions such as low signal-to-noise ratio and long transmission distances. In this study, a modulation classification approach is proposed for underwater wireless optical communication systems that can operate effectively even at low signal-to-noise ratio levels, requires low data, and offers high interpretability. In this context, a three-level discrete wavelet transform was applied to the modulated signals received from the communication channel. The mean, variance, and energy values of the approximation and detail coefficients obtained from the decomposition were used as feature input vectors in a genetic algorithm to evaluate and determine the most suitable feature combinations. It was observed that energy-based features contributed more to classification performance than mean- and variance-based features. Based on this observation, the feature dimension was reduced and a classification accuracy of 0.84 was achieved. The results show that although classification performance decreases at low signal-to-noise ratio levels such as −10 dB and −5 dB, features obtained from the coiflets and biorthogonal wavelet families still provide high accuracy under these challenging conditions.