I find this article to be generally well-written, well-structured, and applies a transferrable methodology to classify stream conditions in ephemeral streams in a single study watershed. The authors support their claims and provide adequate figures to support their argument.

I did find that some of the discussion sections strayed beyond the scope of the study described in the introduction section to discuss other features of the watershed and ephemeral streams more broadly. The paper would be strengthened by focusing on its central contribution.

I did find that a limitation of the study was that it focused on a subset of images from a single site. The methodology was demonstrated and its performance evaluated against predictions from the National Water Model, but statements about its transferability to other locations or are undercut by the limited nature of the data.

 I have minor comments regarding clarity and a few considerations not in the original text but overall find the article a suitable contribution to HESS:

1. Page 7, line 155: What defines “environmental damage”? Tampering? Batteries dying?
2. Page 9, Lines 173-180: The National Water Model (NWM) is trained/calibrated to gage flows, how close is the closest calibration site? In figure A1 looks like it is on the East Fork of Russian River, so not on the stream you are monitoring. Worth pointing out in this section.
3. Figure 5: You need axis labels indicating which axis is predicted and which is observed.
4. Page 9, line 197: Indeed, cropping vegetation may be helpful here – if there is a mediterranean climate, vegetation dynamics and streamflow ephemerality are both highly seasonal the model could learn more from the banks (which could make up more of the image) than the channel where intended.
5. Page 10, line 204: You only labelled 12.8% of the total images you had available – this is acceptable but is a relatively small dataset for training or reporting performance (your testing set is 3.9% of your total image dataset) that will represent the population. This is a limitation of the study, since as you note the lighting can be very different at different times of day/year. Ideally you have a big enough testing set to represent performance at each class during different lighting (and vegetation and channel) conditions.
6. Page 10, line 203-210: Random sampling was used in the selection of images for training/testing, which is acceptable, but this means the performance is only representative of historical conditions coincident with the label dataset. The performance reported in this paper is not representative of model prediction on new unseen imagery. This point is worth noting to make sure the reader knows what the model performance represents.
7. Page 12, lines 247-250: Why were these manual weights selected?
8. Page 14, line 298: Is there a reference for the 0.028 m³ s^-1 threshold for NWM flow? How sensitive are your results to this selection? The selection of the threshold appears arbitrary at the moment.
9. Figure 7: Why are there negative stage values? And why are there purple high water dots in panel 7 when stage is reported negative? Is that supposed to be a diagnostic tool for quality assurance of the stage data, which leads to the record in panel b? The paragraph in the main text where Figure 7 is mentioned does not walk the reader through this. Also in Figure 7 are the stage observations without any dots times where there was no imagery or times where the imagery classification was deemed not high confidence? Consider adding shading to indicate “no imagery available” and another color of dot to indicate “no high confidence prediction” or something similar so the absence of data is clear.
10. Page 28, line 446-448 and Page 29, line 464-465: Is there a citation or the claim of not having enough imagery to train a CNN? The Gupta et al. and Noto et al. studies you cite have about as much imagery as you do. You report ~4,700 images, which is more than at 2 of Gupta et al. 's sites.
11. Section 4.4: This section largely diverges from the central contribution of the study (a methodology to classify images) and into a lot of site-specific information that is largely conjecture about processes and reads as redundant to the prior section (4.3). This section could be eliminated.
12. Page 33, line 604: Is there a citation for the claim that “these efforts have struggled to translate to IRES”? Neither study cited included nonperennial streams.
13. Section 4.6: This section is only loosely connected with the central contribution of the paper (image classification model) and is material that could be included in the introductory material. This section could be eliminated.
14. Conclusion section is missing: It is traditional to summarize the paper’s contribution in a conclusion section, one is missing here.

This manuscript presents a timely, and valuable study that addresses a critical challenge in hydrology: monitoring intermittent rivers and ephemeral streams (IRES). The application of a relatively simple logistic regression model to classify flow states from field camera imagery is both pragmatic and innovative. The methodology is clearly described, the results are robust and convincingly presented, and the discussion thoughtfully places the work in the broader context of IRES monitoring, climate change, and water management. The integration of image classifications for quality control of stage data is a particularly strong and practical contribution. The manuscript is generally well-written and structured. I believe it represents a significant contribution to the field and is a strong candidate for publication after revisions. In the end, I've attached a document with some suggestions that may improve the manuscript.