Time Series Foundation Models and Deep Learning Architectures for Earthquake Temporal and Spatial Nowcasting
, Andrea Donnellan 4 and Lisa Grant Ludwig 5Round 1
Reviewer 1 Report
Comments and Suggestions for Authorsgeohazards-3281538 REVIEW
This paper is a comprehensive and thorough account of the authors’ research work on using deep learning models to improve nowcasting of earthquakes. Based on the calculation of various statistical parameters, the authors evaluate the performance of various models and suggest which ones are better for nowcasting. The paper shows some interesting results that should be of interest to other researchers and should be published. The paper is very well written and organized. I have only a few minor questions and comments.
QUESTIONS AND COMMENTS:
[1] The journals in references 20 and 26 are missing in the references list.
[2] Line 82: What is the meaning of “holschneider2014can” ?
[3] Line 187: Instead of “magnitude is raised to the power 1.5”, do you mean “magnitude is multiplied by the factor 1.5” (see equation 1)?
[4] Line 661: For a better comparison between MSE and MAE, would it not be better to use the RMSE (root mean squared error) instead of the MSE?
[5] In Figure 7, it is not easy to tell apart the various model prediction curves. The colors mix together too much. Perhaps a few blow-ups or zoom-ins of sections of the time series might be useful.
[6] In Table 2, is the last MultiFoundationQuake1 row equivalent to the second MultiFoundationQuake1 row? The same asterisk distribution appears.
[7] Line 764-770: How significant is this improvement? The performance parameters MSE, MAE, and NNSE show numbers that are only slightly different. In fact, many of the performance parameter numbers in the various tables do not differ very much. So, how reliable and general are your conclusions on the performances?
[8] Line 849: However, doesn’t the necessity of selecting appropriate model architectures and pre-training datasets complicate the process of nowcasting? How feasible is the current state-of-the-art? How successful has nowcasting been so far?
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThis manuscript explores data-driven earthquake nowcasting approaches using a range of time series foundation models and deep learning architectures. It is a well-written and engaging study that effectively highlights the importance of capturing both spatial and temporal dependencies in seismic data. The results are promising and demonstrate the potential of the proposed methods. Overall, I believe this work is suitable for publication, with only a few minor points for consideration, as outlined below.
The introduction section could be better organized. The current form is a bit long. It contains a large portion of methodology. You may consider reorganizing it to further improve its readability.
Line 82: it seems there is a typo in your latex (reference).
Line 688: For the undesirable performance of TimeGPT, could you dig a bit deeper on why its performance is poor for your case?
Line 220-246: The creation of the graph structure for GNNCoder, based on the epsilon nearest neighbor algorithm, is interesting. However, more detail on how the epsilon value was determined would improve understanding, as this parameter likely affects the model’s ability to capture relevant spatial relationships. Additionally, discussing potential limitations of this graph structure and any alternative approaches considered would add depth to the methodology.
Lines 586-602: The description of the training process would benefit from additional details on hyperparameter tuning, convergence criteria, and data splitting. Discussing the decisions behind these choices and how they might impact model generalizability could provide readers with insights into the robustness of the model.
Lines 625-638: The manuscript uses the Normalized Nash-Sutcliffe Efficiency (NNSE) metric alongside Mean Squared Error (MSE) and Mean Absolute Error (MAE) to evaluate the models. While NNSE is a good choice for comparing predictions against observed data, a brief discussion on why it is preferred over more commonly used metrics (like MSE) would help readers unfamiliar with NNSE interpret its relevance to the study.
In Figure 7, the visual representation of predictions across six spatial bins illustrates the model’s performance in different locations. However, could it be possible to add a quantitative summary of prediction errors across all bins (e.g., average error or variance)? This may provide a more complete picture of model consistency and reliability across spatial regions.
Lines 247-291: The foundation models used in the study are pre-trained on datasets (e.g., TrafficL, Weather, M4) that may have limited relevance to seismic data. It would be beneficial if the you could clarify the specific temporal or spatial characteristics in these datasets that support their use for earthquake nowcasting. For example, if the Weather dataset captures similar temporal variations, or if the TrafficL dataset provides insights into periodic patterns, these connections should be explicitly outlined. Additionally, any potential mismatches between these datasets and seismic data (e.g., differences in temporal scales or event frequency) should be addressed to justify the transfer learning approach.
Earthquake events are inherently sparse and irregular. This kind a dynamics is challenging for deep learning models that typically require dense time series data. Could you discuss on how sparse or uneven event data was handled, especially since models like transformers and LSTMs can struggle with irregular intervals. You might also consider adding an explanation of any data imputation, aggregation, or resampling techniques applied to standardize the time series for training purposes. This would ensure readers understand how the models manage seismic data’s unique characteristics.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
Comments and Suggestions for AuthorsThanks for taking efforts in revising this manuscript. The work is interesting and valuable. I have no further comments.
