Cold climates pose significant challenges for wind turbines, primarily due to icing that influence electrical energy production. Precise methods are needed to identify and predict ice distribution on blades, enabling enhanced prediction of ice accumulation based on the blade's frequency response. This study uses glass fiber reinforced plastic composite rotor blades equipped with actuators and accelerometers to measure, with a total of 1700 measurements, the response of the blade subjected to icing. Small-scale icing experiments are conducted inside a climate chamber at temperatures ranging from −10∘C to −20∘C with seven ice distribution profiles on the blades. The gathered data is analyzed for the effects of icing on the frequency response of the blades. Optimized artificial neural networks, using fully connected layers and convolutional layers, are proposed to predict the accumulated ice thickness on rotor blades based on the frequency response, with weighted mean absolute percentage errors of 5.1 % and 5.8 %, respectively, and to predict ice volume and ice mass with errors of 5.7 % and 4.9 %, respectively. Overall, this study investigates the effect of icing on the frequency response of composite blades with regard to ice mass and ice location, and proposes a high-performance data-driven method for ice detection and monitoring during operation.