Summary and primary contribution

This study investigates the use of Convolutional Neural Networks (CNNs) and hybrid CNN-Long-Short-Term Memory networks to predict interannual shoreline position. The target observation is a shoreline position at one location derived from 18 years of daily shoreline camera images at Tairua beach, North Island of New Zealand.  The drivers include wave peak period, significant wave height and direction and sea level pressure. The results are compared with a subset of the target observation not used for training or tuning and also two models; ShoreFor and SPADS. Using a systematic search and different measures of fitness the authors conclude that CNNs models have the potential to improve accuracy and reliability over current models.

General comments

The combination of different metrics, graphical results (Taylor diagrams) and grid search and ensemble approach to evaluate shoreline models’ performance is novel and the use of CNNs is yet not well understood, of interest of the coastal engineering community and therefore suitable for publication on the EGUsphere.

The manuscript is well structured and written in a clear fashion and well referenced.

My main concern is on the lack of some methodological important details (see also specific comments) and most critically on the rationale for constraining the prediction to a single location (not shown) to assess the shoreline predictions. This is important, as the authors indicated the oscillatory nature of the shoreline changes (L370). Should the location being close to a nodal point, the same time series of drivers will have produced virtually no changes in the cross-shore location. The ability of the model capturing the shoreline position, simultaneously at different location is not presented and the claimed improvement over ShoreFor and SPADS remains unclear.

Specific comments

The camera system provides images of a section of the Tairua beach but only the cross-shore position at one location has been used as a target but neither the rationale for choosing this location or a map showing the location is presented. 

A short description on the set up used for the ShoreFor and SPADS model need to be included. At present, the manuscript contains very detailed information on how CNNs model has been set-up but no information is provided on the set up of the ShoreFor and SPADS models. To be consistent with authors closing remark (L373), I encourage them to make the model configuration publicly available.

Figure 4 shows both the target and drivers time series, but it is unclear if all time series have the same frequency (daily, hourly, …) and if the shoreline position was corrected for any differences in tidal elevation at the time of the camera image was captured.

The manuscript authored by Gomez-de la Peña et al. investigate the potential of DL algorithms to predict interannual shoreline position derived from camera system observations at Tairua Beach in New Zealand. They investigate the application of Convolutional Neural Networks (CNNs) and hybrid CNN - Long Short-Term Memory networks.

This study offers an innovative approach as an alternative to apply shoreline evolution models, making it a manuscript of significant interest to the coastal engineering scientific community.

Overall, the document is well structured, details each section adequately, and is written in fluent English.

General comments are indicated below and more detailed comments are in the attached document:

• While the manuscript provides sufficient details about the data and study site, it would greatly benefit from a figure that situates the reader in the study area and highlights the described elements. To enhance clarity, it is advisable to include a figure that portrays the study area's location, outlining the video camera system's position, the monitored coastline section, and the wave point utilized for forcing, among other pertinent features.
• The shoreline position time series depicted in Montaño et al. (2020) displays more fluctuations compared to the one presented in Figure 4 of this manuscript. It would be valuable to clarify whether the time series corresponds to raw data or processed data, such as a moving average. To enhance clarity, presenting the raw data as points rather than a continuous line in Figure 4 would enable readers to identify any potential gaps in the measurements.
• In the manuscript, it is recommended to specify the two distinct meanings of the term 'memory': one as memory cells or memory blocks in DL algorithms and the other as the 'memory decay function' employed in the ShoreFor model.
• Has the performance of the suggested approach been assessed considering different calibration period extensions? Is there a specific minimum timeframe or minimum quantity of data necessary for the application of this methodology?