Deep learning applications for hydrological forecasting have become increasingly common over the past two decades. Among various deep learning models, the hybridization of the Convolutional Neural Networks (CNNs) with the Recurrent Neural Networks (RNNs), particularly the Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks, has gained popularity due to their capability of capturing both spatial and temporal features. This study investigates how different kernel sizes and input channels have an impact on the performance of the CNN, CNN-LSTM and CNN-GRU in dam water level forecasting at the Klang Gates Dam, Malaysia. The results indicated that the increase in kernel size has no impact on the performance of the CNN because the output layer was a fully connected feed forward layer. However, the LSTM and GRU layers, which process the output of the CNN layer sequentially, performed better when the kernel size was smaller due to the finer detail extracted from the CNN layer. Due to the different nature of input features, the input channels of the CNN layer were split into water level data and multivariate data input channels. Both hybrid models showed that decreasing the kernel size of the water level data input channel and increasing the kernel size of the multivariate data input channel, have improved the model performance. With the splitting of the input channels of the models, the performance of the CNN, CNN-LSTM and CNN-GRU models, with their respective best kernel size configurations, showed improvements of 13.57%, 14.39% and 15.87%, respectively. Hence, the splitting of input channels with more model configurations are recommended for further studies.