I would like to thanks the editor and congratulate with the authors for giving me the opportunity to review this very interesting paper.

I find the research in its conception and in the methodologies applied extremely interesting and with many novel and innovative points. Moreover, its significance in the field of debris flow monitoring is very high.

I just have some minor comments and questions:



1- It is my understanding that using hillshade representation, only 2 dimensions can be analyzed. Is there any possible improvement to obtain also the vertical dimension, which could be very important to be estimated?

2- Would this methodology be implemented also to forecast trajectories of boulders and woods?

3-Which are the error bounds of the different size materials in terms of velocities using Sort and Bot Sort?

4- Can this method be extended to differentiate submerged vs. surface-level boulders?



As a mere suggestion on the arrangement of the paper, I would split section 2 into 2.1 Geological and geomorphological setting (providing more info about the catchment area) and 2.2 about the monitoring set up. I find it a little bit confusing as it is.



All the best

 I would like to thank the opportunity for reviewing this paper, which presents a novel and beneficial approach to automating debris-flow feature detection using LiDAR-derived hillshade images and deep learning. In general, the work is well-structured, and offers valuable insights into granular dynamics and longitudinal sorting. However, several issues are still needed to strengthen before the publication in NHESS.

1. Please more explicitly describe the algorithms and parameters used for coordinate transformation, wobble correction, and artifact removal to ensure reproducibility.

2. Authors should briefly explain why these six specific models (YOLO variants, RT-DETR, etc.) were chosen over others, especially given their similar performance.

3. Could you discuss why BoT-SORT was ultimately preferred over SORT in more detail, especially regarding high-velocity object handling?

4. Add confidence intervals or p-values to velocity and size comparisons to strengthen quantitative claims.

5. How is the loss of 3D information in hillshade projection? And its effect on vertical velocity underestimation?