RLNformer: A Rainfall Levels Nowcasting Model Based on Conv1D_Transformer for the Northern Xinjiang Area of China

Round 1

Reviewer 1 Report

In this paper, the authors introduce a novel data-driven rainfall nowcasting approach using transformer-like neural networks. They implemented their method to forecast four categorized rainfall patterns in the Xinjian region of China, highlighting the superiority of their proposed approach compared to traditional statistical methods.

 

General comments:

The topic of this paper is suitable for this journal. However, there are several areas of improvement to maximize the paper's potential. I recommend the editor to reconsider this paper after major revisions.

 

Major points:

L25-148: The introduction can be enhanced in multiple ways:

1. The authors should avoid delving into topics not directly pertinent to precipitation nowcasting and concentrate more on the context of precipitation nowcasting itself. Notably, the mention of ecology (L30) seems irrelevant since nowcasting primarily targets 0-3-hour predictions, mainly concerning heavy rain and flood predictions. Extended lead time predictions are more relevant for discussions related to water scarcity and ecology.
2. The criticism directed towards the physical method seems out of place. Most operational centers rely on statistical models for nowcasting, as physics-based atmospheric models become more effective only with an increase in lead time. The latter isn't employed for nowcasting not because they're computationally intensive but due to their comparative ineffectiveness against statistical models. The authors do not need to criticize the physics-based models since they are not used.
3. There seems to be a discrepancy in the paper's understanding of precipitation nowcasting. It usually implies the prediction of spatial rainfall distribution based on meteorological radar-based observations, while the authors seem to use point-scale weather station observations. This could cause confusion for many potential readers, and I recommend the authors to clarify this point.
4. The authors' scientific questions are unclear. While they referenced and detailed numerous previous machine learning-based precipitation nowcasting studies, they failed to highlight the limitations inherent in these studies that their paper intends to address. They underscored the three contributions in the introduction's concluding section. The first point seems tenuous. Merely choosing Northern Xinjiang cannot stand as a distinct scientific contribution in the context of this paper. The second point does bear the potential of being a genuine scientific contribution. However, it remains unclear how the RLNformer could overcome the shortcomings of the existing methods the authors cited. I must admit that I am not good at neural networks, but given the myriad possible neural network architectures, merely proposing a new one doesn't inherently denote a significant advancement. I recommend the authors to pinpoint the deficiencies of previous methods and articulate how their proposed technique addresses them. I could not agree with the third contribution, given that numerous studies already evaluate prediction methods, taking into account unbalanced samples.

 

 

L277: The process of feature selection, from 90 to 80, seems minimal. Is this reduction genuinely essential? Is it not feasible to utilize all 90 features for model construction? Could a higher MIC score cutoff threshold allow for more significant feature reduction?

 

 

Minor points:

L285-L293: This section and Figure 3 seem redundant. They offer little of interest to hydrologists. Such content might be appropriate for a computer science journal, which is not the case here.

 

Figure 4: The figure is unclear due to its brief caption. Clarifications are needed on what the various box colors represent and the meaning of the grey arrow.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

In this study, six meteorological stations in the northern Xinjiang area were selected as the research area. Due to the high volatility of rainfall in this area, the rainfall was divided into four levels, namely "no rain", "light rain", "moderate rain", and "heavy rain and above" for rainfall levels prediction. In order to improve the prediction performance, this study proposed a rainfall levels nowcasting model based on Conv1D_Transformer (RLNformer). Firstly, use the maximum information coefficient (MIC) method for feature selection and sliding the data, that is, using the data of the first 24 hours to predict the rainfall levels in the next 3 hours; Then, the Conv1D layer is used to replace the word embedding layer of the Transformer, enabling it to extract the relationships between features of time series data and allowing multi-head attention to better capture contextual information in the input sequence; Additionally, a normalization layer is placed before the multi-head attention layer to ensure that the input data has an appropriate scale and normalization, thereby reducing the sensitivity of the model to the distribution of input data and helping to improve model performance. To verify the effectiveness and generalization of the proposed model, the same experiments were conducted on the Indian public dataset, and seven models were selected as benchmark models. Compared with the benchmark models, RLNformer achieved the highest accuracy on both datasets, which were 96.41% and 88.95% respectively. It also has higher accuracy in prediction of each category, especially the minority category, which has certain reference significance and practical value. How the present study differs from the previous literature? Many methods have been proposed and each study claim outperforms the previous one.Section 2 title is very confusing. Please revise it.Why minimum-maximum normalization is used? Why not any other?Please explain equation 6 more.Please report computational time of the methods compared in this study.In data analysis section, why hyperparameters are manually tuned?A recent innovation in the high frequency data is the functional data analysis, which can be used in the current study. If not, please mention it as the future work. See for example,https://www.mdpi.com/2227-7390/10/22/4279https://link.springer.com/article/10.1007/s42452-020-2238-xhttps://iwaponline.com/jwcc/article/11/4/1748/70338/Functional-data-analysis-of-models-for-predicting

A minor check is required to remove typos.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

In this study, six meteorological stations in the northern Xinjiang area were selected as the research area. Due to the high volatility of rainfall in this area, the rainfall was divided into four levels, namely "no rain", "light rain", "moderate rain", and "heavy rain and above" for rainfall levels prediction. In order to improve the prediction performance, this study proposed a rainfall levels nowcasting model based on Conv1D_Transformer (RLNformer). Firstly, use the maximum information coefficient (MIC) method for feature selection and sliding the data, that is, using the data of the first 24 hours to predict the rainfall levels in the next 3 hours; Then, the Conv1D layer is used to replace the word embedding layer of the Transformer, enabling it to extract the relationships between features of time series data and allowing multi-head attention to better capture contextual information in the input sequence; Additionally, a normalization layer is placed before the multi-head attention layer to ensure that the input data has an appropriate scale and normalization, thereby reducing the sensitivity of the model to the distribution of input data and helping to improve model performance. To verify the effectiveness and generalization of the proposed model, the same experiments were conducted on the Indian public dataset, and seven models were selected as benchmark models. Compared with the benchmark models, RLNformer achieved the highest accuracy on both datasets, which were 96.41% and 88.95% respectively. It also has higher accuracy in prediction of each category, especially the minority category, which has certain reference significance and practical value.

 

The authors addressed all my previous concerns and I am happy to recommend it to publish in the present form.