Publisher’s note: the content of this comment was removed on 5 November 2025 since the comment was posted by mistake.

This comment was prepared as part of MSc course work at Wageningen University under supervision of Prof Wouter Peters. They were uploaded as a comment as they were regarded to be of good quality, and likely helpful to the authors and editor in the review process.

This study examines how aerosol regimes affect energy and carbon fluxes in a pristine central Amazon forest. Using 2016–2022 meteorological and flux data from the Amazon Tall Tower Observatory (ATTO) and AOD (500 nm) from AERONET, it tests whether aerosol loading alters latent heat (LE), net radiation (Rn), and CO₂ fluxes (FCO₂). The study focuses on the dry season (August–November), when biomass burning elevates aerosol concentrations across the southern Amazon Basin. The authors define two aerosol regimes: clean (AOD < 0.13) and polluted (AOD > 0.40), consistent with previous studies such as Steiner et al. (2013) and Ross Herbert & Stier (2023). This threshold-based approach, derived from data percentiles, provides a simple yet robust framework for distinguishing contrasting aerosol loading conditions. Their analysis focuses on the 10:00–14:00 LT period to examine energy partitioning under contrasting aerosol regimes. Authors interestingly present, VPD vs Temperature (Figure 4), a combination of variables that I have not encountered in other studies reviewed during this process. It is particularly valuable, as it effectively illustrates—through the observed variables of VPD and temperature—the realistic delay caused by reduced shortwave incoming radiation (SWin) during polluted periods. They report a delay in the rise of temperature and VPD under polluted conditions, highlighting the moderating effect of aerosols. They conclude by confirming the well-documented finding that, paradoxically, CO₂ uptake increases under polluted conditions—by about 57.7% in this case—due to the diffuse radiation effect, where scattered sunlight enhances photosynthesis within shaded canopy layers. This result is in strong agreement with previous studies on Amazonian aerosol dynamics, particularly Rodrigues et al. (2024) and Cirino et al. (2014), which similarly observed elevated carbon uptake under high-AOD conditions. The study concludes by emphasizing the nonlinear and complex interactions between AOD and surface fluxes, demonstrated through MANCOVA and Random Forest Model analyses, underlining however the need for further investigation.

Remarks on several aspects:

(1) Midday Averaging

The authors assess the effects of aerosols on surface energy and carbon fluxes by averaging 30-minute flux measurements over the 10:00–14:00 LT period and then calculating percentage reductions between clean and polluted aerosol regimes. This time window is identified as representing the period of strongest radiative and convective activity (line 173). However, the diurnal cycle plots (Figs. 5 and 6) reveal uneven flux patterns, with noticeable uninvestigated areas ( Figure 5 & 6 “white spaces”, outside 10:00–14:00 LT window) within both the clean and polluted regimes. As a starting point, Figure 4 clearly shows a delay in the increase of temperature and vapor pressure deficit (VPD). Because natural processes evolve non-uniformly throughout the day, using a short and non-equidistant time subset which may bias the calculated percentage reductions and misrepresent the actual aerosol influence.  The paper’s methodology follows Steiner et al. (2013), who also analyzed fluxes over the 10:00–14:00 LT period and compared similar aerosol optical depth (AOD) regimes (AOD < 0.3 vs. > 0.5). However, within the text, fluxes reductions’ comparisons are made with studies that employed different approaches to assess aerosol-load effects. For example, Rodrigues et al. (2024) and Cirino et al. (2014) estimated flux reductions under specific irradiance conditions, distinguishing Solar Zenith Angle (SZA) zones and thereby incorporating the time-of-day variability, rather than relying on a fixed midday average. A closer examination of Figures 5 and 6, which depict sensible, latent, and ground heat fluxes, reveals an interesting but unexamined pattern during i.e. the morning transition (06:00–10:00 LT). Both H and G occasionally exceed their respective values under polluted conditions, while the consistent dominance of the clean regime in LE appears to be underestimated. Early-morning CO₂ uptake (Figure 6) also exhibits a more dynamic behavior, with pronounced transitions between clean and polluted regimes. To better capture the full evolution of the phenomena and associated fluxes, the authors could integrate the area under the fluxes’ curves over the 06:00–17:00 LT period and compare the resulting averages between the clean and polluted aerosol regimes. Alternatively, if there is sufficient data outside the window 10:00-14:00 LT the authors could consider reporting morning (06:00-10:00 LT) and afternoon sub-period (14:00-17:00 LT) averages separately to capture diurnal variability better. Analyzing relative irradiance would require substantially more methodological development and investigation by the authors; therefore, it is not recommended.

(2) Gaps Manipulation

The authors state that their initial dataset comprised 10,890 half-hourly observations (line 87), which, after several filtering steps, was reduced to 523 rows—of which only 370 belong to the dry season (lines 94–96). However, the paper does not clarify how these 10,890 records were originally obtained. Figure 2 further raises questions about data representativeness and statistical treatment: the monthly boxplots show means much higher than medians, indicating positive skewness, while the number of valid data points per month is not reported. The data filtering process is clearly described, resulting in 523 rows of 30-minute averaged meteorological, flux, and AOD values. However, the dataset distribution across years is highly uneven, as also noted by the authors (line 97: “The distribution…effects of aerosol”). Specifically, years contributing less than 5 % of the total dataset are treated equivalently to years such as 2020 and 2022 (42,4% and 29,2% data coverage respectively), despite potentially different atmospheric and surface conditions. This raises concerns regarding the robustness of the study’s conclusions. Evidentially, no quantitative assessment of data representativeness or uncertainty is provided. Similar studies (e.g., Schmitt et al., 2023) have explicitly visualized monthly data availability and included “fraction of missing data”. Moreover, the extremely low number of data rows for certain years warrants further examination, as such sparse temporal coverage could substantially affect the robustness of the Random Forest Model (RFM) used later in the statistical analysis. Limited data availability may lead to overfitting, biased feature importance when training and validation subsets are unevenly represented. It is recommended that the authors include the fraction of valid rows per month, which could be directly incorporated into Figure 2. Furthermore, the manuscript should clearly describe the origin of the initial 10,890 observations—specifying the time period covered, sampling frequency, and measured variables—to better contextualize the subsequent data filtering process.

(3) Statistical Analysis

The study explores the relationship between aerosol optical depth (AOD) and surface fluxes (Rn, H, LE, FCO₂) implementing Spearman correlations, multivariate MANCOVA testing assessed by Pillai test and a Random Forest Model (RFM) to quantify nonlinear dependencies and variable importance. However, several methodological lack in processes or data-handling limitations seem to weaken the robustness of the conclusions. The manuscript provides a general introduction to the application of Pillai’s test and outlines the advantages of using the Random Forest Model (RFM) to investigate nonlinear and complex interactions between variables and systems. However, it remains unclear to what extent these principles—particularly in the case of RFM—have been appropriately implemented and demonstrated in the study. In comparable RFM environmental works, such as Miao et al. (2018) and Zhang et al. (2023), linear correlation analyses were explicitly conducted to assess collinearity among key variables by providing comprehensive correlation matrices, providing direct linear insights. In contrast, Rocha et al. (2025) only briefly mention in line 272 that “the statistical relationships show low intensity or no statistical significance,” without offering supporting analyses or graphical evidence. Furthermore, while Miao et al. (2018) thoroughly examined their multivariate equations and reported the statistical significance of their models and variables, Rocha et al. (2025) limit the discussion to the significance of Pillai’s test (line 275), suggesting the absence of linear interactions without presenting sufficient analytical support or methodological transparency. Another major concern is data volume, as mentioned in major argument 2. Miao et al. (2018) utilized approximately 7,000 samples, and Zhang et al. (2023) worked with about 60,000 samples. In contrast, Rocha et al. (2025) rely on only 370 rows of data for the dry period, which raises serious concerns about potential overfitting of the RFM. Moreover, although the manuscript mentions a cross-validation approach in Table 3, it does not specify the technique used or report its results. Finally, the model assessment presented in Table 3 appears inadequate and leaves substantial uncertainty regarding the RFM’s reliability. In the referenced studies, Miao et al. (2018) implemented multiple factor matrices, and Zhang et al. (2023) validated their models through scatter density plots and strong statistical metrics across training and testing datasets, including mean absolute error (MAE) and percentage variation analyses. Rocha et al. 2025 attempt to employ a RFM to capture the nonlinear influence of aerosols on surface fluxes. However, this approach lacks sufficient methodological justification and statistical robustness. The authors do not provide any evidence of cross-validation or other procedures to assess model generalisability. Furthermore, the dataset used for training—only 370 observations—is several orders of magnitude smaller than what is typically required for stable Random Forest performance, raising serious concerns about overfitting and the reliability of the reported metrics. Consequently, the predictive results presented in Table 3 should be interpreted with caution, as their statistical validity is uncertain. Given the limited dataset, the application of the Random Forest Model (RFM) in this study does not appear to add substantial value to the results or discussion. With such a small sample size, the model’s capacity to generalise is minimal, and its predictive performance cannot be reliably validated. Moreover, the manuscript provides no detailed explanation of the model evaluation or validation procedures, which further undermines confidence in the reported outcomes. To strengthen the analysis, I suggest replacing or complementing the RFM with a correlation matrix to explicitly reveal potential collinearity among variables, particularly regarding the influence of AOD (as in Table 3). Additionally, presenting multivariate regression equations and reporting their levels of statistical significance would offer a clearer and more interpretable understanding of how other environmental factors interacts with AOD. Such an approach could also serve as a solid foundation for future studies investigating aerosol impacts on surface fluxes under polluted regimes.

Minor arguments and typos:

Minor issue 1: Several sentences are poorly structured or ambiguous, leading to confusion or misinterpretation. Examples include lines 74–75, 97, 99–101, 112–113, 134, and 247–248, as well as the descriptions for Figures, especially 2 and 4, where I suggest rephrasing or clarifying.

Minor issue 2: Several statements lack adequate justification or references, I suggest further elaboration on the statements:

o Line 114: The use of a fourth-order polynomial is mentioned but not explained or visualized.

o Lines 135–136: Require citation or elaboration.

o Lines 220–222: Could be expanded with a brief example of the described method.

Minor Issue 3: Some methodological descriptions (e.g., line 12 in the Abstract; lines 87–90 on data filtering; lines 137–144 on the RFM methodology) could be condensed, as they do not add substantial value to the manuscript.

Minor Issue 4: GPP is mentioned in the Abstract and Conclusion but is neither discussed nor analyzed in the main text.

Minor Issue 5: The manuscript refers to two towers at the ATTO site but does not specify which tower’s data are used in the analyses and figures.

P1, line 12: The last sentence of the Abstract adds no clear value to the manuscript and could be removed.P3, line 81: Change LiCor to LI-COR for correct company citation.

P5, line 112: The text states that hourly averages are used, while figures show 30-minute values—this inconsistency should be corrected.

P14, Table 3 description: Typo — change FCO to FCO₂.

P15, line 312: Typo — change aerossol to aerosol.

References:

A. L. Steiner, D. Mermelstein, S. J. Cheng, T. E. Twine, and A. Oliphant, “Observed impact of atmospheric aerosols on the surface energy budget,” Earth Interactions, vol. 17, no. 14, pp. 1-22, 2013, doi: 10.1175/2013EI000523.1.

G. G. Cirino, R. A. F. Souza, D. K. Adams, and P. Artaxo, “The effect of atmospheric aerosol particles and clouds on net ecosystem exchange in the Amazon,” Atmospheric Chemistry and Physics, vol. 14, pp. 6523-6543, 2014, doi: 10.5194/acp-14-6523-2014.

R. Palácios and F. G. Morais, “ENSO effects on the relationship between aerosols and evapotranspiration in the south of the Amazon biome,” Environmental Research, vol. 250, p. 118516, 2024, doi: 10.1016/j.envres.2024.118516.S. Rodrigues, G. Cirino, D. Moreira, A. Pozzer, R. Palácios, S-C. Lee, B. Imbiriba, J. Nogueira, M. I. Vitorino, and G. Vourlitis, “Enhanced net CO₂ exchange of a semideciduous forest in the southern Amazon due to diffuse radiation from biomass burning,” Biogeosciences, vol. 21, pp. 843-868, 2024, doi: 10.5194/bg-21-843-2024.

A. U. Schmitt, F. Ament, A. C. de Araújo, M. Sá, and P. Teixeira, “Modeling atmosphere–land interactions at a rainforest site – a case study using Amazon Tall Tower Observatory (ATTO) measurements and reanalysis data,” Atmospheric Chemistry and Physics, vol. 23, pp. 9323-9346, Aug. 24 2023. doi:10.5194/acp-23-9323-2023.

S. Miao, X. Zhang, Y. Han, W. Sun, C. Liu, and S. Yin, “Random Forest Algorithm for the Relationship between Negative Air Ions and Environmental Factors in an Urban Park,” Atmosphere, vol. 9, no. 12, Art. no. 463, 2018. doi: 10.3390/atmos9120463.

J. Zhang, Z. Duan, S. Zhou, Y. Li, and Z. Gao, “Gap filling of turbulent heat fluxes over rice–wheat rotation croplands using the Random Forest model,” Atmos. Meas. Tech., vol. 16, pp. 2197–2207, 2023. doi: 10.5194/amt-16-2197-2023.