Thanks for your revision. The adjustment on the paper organization after discussion phase of HESS makes the paper more readable and enables the novelty to be more explicit. Here are some points that might be valuable for further improvement:
(a) Add some experiments
The necessity of introducing transfer learning into the study has not been fully explored in the paper, which would undermine the scientific significance of your research. It is necessary to set up a control group without using transfer learning method to illustrate the significance of transfer learning. And the comparison between this experiment and the two experiments using transfer learning should be illustrated in table or figure. Through the comparison, the scientific significance of the transfer learning in this task will be further clarified.
(b) The focus of the introduction
Keeping the description on other types of water level observation and hydrological uses of river cameras in the introduction part is surely no problem, it is just the matter of length. Since the novelty of the work is the modification on the existing deep learning methods rather than to propose a new model to replace the traditional observation methods, it is still recommended that you emphasize more on the existing computer vision methods for water segmentation. From line 58 to line 65, you can add more reference including research on histogram analysis and machine learning methods for water segmentation. The difference and commonality of these methods are also worth introducing. These computer vision methods are the basis of your work, they deserve more description.
(c) The details about the CNN model
A CNN based model is typically composed of three types of layers, which are convolutional layer, pooling layer and fully-connected layer. However, in Section 2.1.2, only the computational process in convolutional layer is introduced. Are the other two layers used in the model in your work? If so, please supplement the introduction of them, and revise Figure 2 according to the added introduction. This will also help readers to understand the main difference between the two transfer learning strategies in Section 2.2.3 and the scientific significance of the work on comparing different strategies. Additionally, it is also necessary to simply mention the activation function between different layers in CNN model, e.g., ReLu or Softmax.
(d) The setup of the experiments
In Section 3, some key information about model training should be supplemented. Firstly, the setup of hyperparameters is supposed to be added, including learning rate, training epoch, batch size as well as the optimizer for training. Simultaneously, the introduction of learning rate setting would help to understand the meaning of fine-tuning if compared with the learning rate in original training with large dataset. Secondly, the machine learning library used in this study needs to be illustrated, Pytorch, Tensorflow or Keras? Thirdly, which loss function is used for supervising the training should be illustrated, cross entropy loss, MSE loss or others? Fourthly, could you please give some comparison of two training strategies on convergence speed?
Above we provide some general notes about new experiments, introduction, model structure and experimental setup. You could take them into consideration for revision.