This study evaluates the potential of low-cost UAS data for estimating biomass dry matter weight (DM) and nitrogen (N) concentration in Alpine grasslands. A combination of 2 UAS (different sensors and UAV), multiple predictors (spectral and ground based) and algorithms (regression and machine learning) are tested and compared showing moderate performance in assessing DM and poor performance in assessing N, with machine learning algorithms better performing than regression.

The topic is of interest and relevant for the Journal and the manuscript is clear and, overall, well written. My main concern is on two main points:

• The study is entirely based on a single field campaign, therefore during one specific stage of grasslands phenological development (presumably early season). This is an important limitation which is not discussed. I would recommend to more explicitly highlight this and discuss the implications.
• The study in the introduction suggests that UAS can bring significant advantages as compared to satellite or airborne data in mountain regions. However, I would argue they have also significant limitations, especially when low-cost sensors are used. By looking at the study results, I have the impression (perhaps wrong) that issues related to data acquisition and quality (e.g. long acquisition time, calibration and incident radiation measurement) might play an important role in explaining the relatively poor model performances. However, this is not mentioned or discussed in a clear way (there are some points, but not really a section discussing the challenges in UAS data acquisition and quality). I would find valuable to see some more discussions on the issues that might be related to UAS use (even better if supported by some analyses).

Minor comments

Introduction:

• There is a bit of mix between Alpine/pre Alpine etc, while most statements are valid for both. Perhaps if the area falls within the Alpine space (geographic region) there is no need to specify Pre. Alpine is sufficient and more details about the sites are given in the methods.
• Perhaps ‘often long’ is unnecessary
• The sentence is unclear
• 86-89 This might be truth for specific cases, but it is important to keep in mind the limitations of sensors technologies onboards UAS.
• Here and elsewhere it is mentioned canopy height data were not ‘available’ without explanation. I would suggest avoiding that, as this is explained in the methods. Otherwise, a short justification should be added here too.

Materials and methods

• There is no mentioning of the phenological stage of vegetation during the field campaign. This is an important factor.
• 9.50 to 16.30 is a quite long interval with expected variations in solar angle and, in mountain regions, shadows and possibly cloudiness. This could be quite a relevant factor affecting the data acquisition.
• Is there any indication of the geolocation accuracy?
• An area of 3x3 pixels seem very small considering geolocation errors. Assuming the plot should be somehow representative of a wider area, would not be more prudent to have a larger window?

Results

• It is somewhat surprising the NIR does not follow DM or height, as this should be rather straightforward (unless for very small range). Is there any factor related to the acquisition that might be causing this issue?
• The doubt of a strong influence of acquisition factors is also supported by the very poor performance of regression as compared to machine learning and the improved performances on DM including VI in the validation. Also the important role of ground canopy height may suggest that as this variable is clearly not affected by the UAS data acquisition. I would suggest to run some tests and eventually add some considerations in the discussions.
• The word ‘notably’ is very often repeated. Sometime is a bit redundant
• Figure 7. It is a bit strange to see many points along a line (i.e. same N content) in Fig.7b. Is it correct?

This manuscript provided by Schucknecht et al. presents a study of estimating pre-Alpine grassland aboveground dry biomass and plant nitrogen concentration using Unmanned Aerial Systems with two low-cost multispectral sensors. This study tested three statistical models including linear models, random forest, and gradient boosting machines to predict dry biomass and plant nitrogen concentration from UAS multispectral imagery. Three science questions on (1) whether spectral information of UAS sensors is enough for mapping dry matter and plant nitrogen, (2) the need for machine learning hyper-parameter tunning, (3) and model performance with different sensors, statistical models, and inputs were addressed. Results show that the two UAS multispectral data sets can achieve moderate performance to quantify grassland dry biomass and plant nitrogen concentration. Specifically, the best performance of quantifying dry biomass came from the combination of random forest, all predictors, and REM sensor data. The best model for plant nitrogen concentration was achieved by using random forest, all predictors, and SEQ sensor data. Considering the rapid development of UAS remote sensing for mapping vegetation traits, this study is necessary and interesting. The manuscript is well-structured and easy to follow. However, the current manuscript has several issues with experimental designs and the result interpretation. I suggest a major revision for the current format. Here are some comments that may be helpful to improve the manuscript.

Main issue:

1. The motivation for comparing these two UAS sensors is not clear. Are these two types of sensors are popularly used in UAS remote sensing studies? How the findings from the two sensor comparison are relevant to other studies and the UAS remote sensing community? Overall, SEQ and REM sensors are very similar. These two sensors have similar pixel resolution, similar wavelengths in green (550/560 nm), red (660/668 nm), and red edge (735/717 nm). Furthermore, the manuscript pointed out that SEQ performed well for predicting plant nitrogen concentration, while REM had a better performance for predicting dry biomass. However, it is not clear why these two sensors had such different performances in the current manuscript. The analysis and explanation for sensor performance on dry biomass and nitrogen predictions need to be strengthened.

In Table 2, you labeled 790nm as near infrared. However, we usually refer to 700-800nm as red edge, while wavelengths beyond 800nm as near infrared. From the soil-vegetation radiative transfer modeling view, red edge wavelengths are vital for vegetation chlorophyll content and nitrogen content retrieval. The near infrared is more sensitive to the vegetation canopy structure such as leaf area index and total biomass. From my interpretation, SEQ has two red edge bands and could potentially get better results for nitrogen concentration retrieval, but not dry biomass as lacking information in near infrared. Meanwhile, REM has information on near infrared which is good for biomass retrieval.

2. The motivation for selecting Gradient Boosting Machines and Random Forest is also not clear. Why not other more popular machine learning or statistical approaches, such as partial least-squares regression, LASSO, Ridge, or Neural Networks?

The purpose of applying machine learning algorithms is not only to achieve good model predictive performance. Many machine learning algorithms like random forest can help to identify the relative importance of each feature input. This feature importance analysis is very necessary to understand the relationship between feature inputs and the predicted variables. However, such analysis is missing in this study. I strongly recommend further feature importance analysis to identify scientific linkage among input variables and the predicted variable to strengthen the manuscript result interpretation.

3.  The UAS multispectral data were collected from one single flight in each site. How robustness of these results across different growth stages and dates is uncertain?

4.  Machine learning parameter tunning is a very necessary and common step to implement model training. However, this manuscript highlights the hyper-parameter tunning as one major research question. The innovations of this study need to be strengthened.

Minor issues:

1. There are many abbreviations in Figure 2. The caption should add explanations of these abbreviations for readers.
2. The reflectance values in Figure 4 look quite different from the two sensors. Do you have ground reflectance collection to validate your reflectance?
3. The manuscript mentioned that mountain regions have frequent cloud occurrences to argue the weakness of Copernicus satellite missions. However, UAS data collection under cloudy environment also has data quality issues. The manuscript may need to discuss such potential issues and mitigation strategies.
4. Most parts of the manuscript used nitrogen concentration. However, Figure 6 used nitrogen content in the (c) and (d) subplots.
5. The same issue of nitrogen concentration on Figure 7.
6. Figure 8 (d) has clear shadows. The reflectance from these shadows needs to be either corrected to real surface reflectance to quantify vegetation traits or simply removed. I don’t think the current estimates for areas in tree shadows are right.
7. The figure panel design of Figure 8 is strange. We normally put RGB into the first subplot. You have paired maps for DM and N. These paired subplots could be in one row.