Please find attached the document revision.

Great job, interesting work. A few comments and suggested changes:

Some food for thought... If I'm using gamma logs as input parameters to predict formation, for instance, and I also have VNIR-SWIR-MWIR-LWIR data for the same holes and find that it can accurately predict gamma values, why not directly predict lithology? Same logic for sonic logs... If the reason for acquiring sonic logs is to log porosity/permeability, why not directly predict that rather than travel times? This is just a thought exercise, to spur you to think about end-user applications, it in no way invalidates your work. 

Section 3.3 (Data co-registration). Downhole geophysical tools typically start measuring distance from surface at 0 and measure in a linear fashion downhole based on how much line has been unspooled. Really they have the best depth registration of virtually all the drillhole analysis methods, including core scanning. Core scanning hardware typically registers depth between driller blocks or on a per-box basis. Not as accurate, and depending on the circumstances may be significantly different from the depths provided by wireline geophysical tools. Section 3.3 does not address this issue of co-registration. Or perhaps there is an underlying assumption that the scanned data depths are accurate and correspond to the geophysical depths? Either way this should be addressed or at least acknowledged. It's one of the greatest barriers to performing ML workflows on drillhole data.

Section 3.5 (Data balancing): You've used HBSCAN to cluster the data and further identify noise (class -1) which you've removed from the dataset. You've mentioned resiliency to hyperparameter selection--- it is actually very well established that HDBSCAN outcomes are highly sensitive to hyperparameter selection, particularly the distance metric, min_samples, and min_clusters parameters. sklearn can automate hyperparameter selection using Randomized Search Cross Validation, which seeks to optimize the validity index for iterated hyperparameters, otherwise hyperparameter tuning is highly manual and hugely impacts the number of clusters and cluster distributions returned. I'd strongly suggest addressing this.

In the 3rd paragraph of section 3.5 you describe using stratified sampling; what was the category that you stratified? The drillhole ids? The HDBSCAN clusters? Did you stratify to bins of a petrophysical parameter? Please clarify.

In general I think you've done a great job of transmitting complex themes accessibly. The workflow described, aside from the above points, is clear and seems robust.

The manuscript presents well-structured research on predicting petrophysical properties using hyperspectral drill core data, leveraging deep learning techniques. The methodology is scientifically sound, and the application of a multi-headed 1D CNN, data preprocessing techniques and Shapley analysis for interpretability are strong aspects of the study.

However, there are some areas where improvements or clarifications would enhance the paper, particularly in terms of abstract impact, workflow clarity, model generalisation and uncertainty quantification. These concerns have been outlined in the attached file.