This paper presents a thorough update on previous work, highlighting several important improvements in the Retina algorithm. The topic is relevant, and the paper is both well-structured and clearly written. As the authors emphasize, the most significant advancement over their earlier work is the integration of satellite data into the data assimilation scheme, which apparently would only have an added value when less than 5 monitoring stations are available in a specific city. 

Specific comments:

1. Since the main novelty of the presented methodology lies in the integration of satellite data (TROPOMI) for urban NO₂ modeling, a longer validation period would be highly valuable. The current one-month evaluation period, may not fully capture the seasonal variability in satellite retrieval quality, atmospheric dynamics, and emissions. Statistical performance metrics may therefore exhibit seasonal dependence, potentially leading to less robust conclusions regarding the added value of satellite data. 
2. The method used to estimate background NO₂ concentrations via a line integral over the municipal perimeter raises several questions: (i) Does $\vec{e}_v$ represent the local wind direction? If so, at what altitude or vertical level is the wind taken from? (ii) Eq. 1 resembles a mass conservation approach, but it lacks a temporal term—how is accumulation of pollutants within the domain accounted for? (iii) Since the method depends on integrating along the perimeter, does this imply that the background concentration depends on the chosen perimeter? (iv) In the special case where $\vec{e}_v \cdot \vec{n} > 0$ for the entire perimeter (i.e. all wind is outflow), the integral appears ill-posed. How is this handled in the analysis? (v) Is this a novel approach? If so, could the authors justify its use and provide a comparison with background concentrations derived from station data within the domain?
3. The manuscript suggests that the added value of TROPOMI measurements becomes negligible when data from 5 or more stations are available. However, this rule of thumb may not be sufficiently robust, as it oversimplifies the issue. Other factors (such as city size, NO₂ concentration levels, local meteorological conditions, …) can significantly influence this threshold. 
4. The manuscript states that traffic flow between counting locations is estimated using inverse-distance weighting interpolation, applied separately for highways and primary roads. However, since traffic volumes can vary significantly over short distances, especially in complex urban settings, this method might lead to unrealistic flow patterns. Could the authors justify the use of this interpolation approach and provide information on how its performance was assessed? Specifically, has any cross-validation been performed (e.g., removing some sensors and comparing interpolated vs. observed counts)?
5. Pg.  27 line 518 : The manuscript compares the Retina model’s performance in Madrid with that of Kim et al. (2021), who trained a model using TROPOMI and 340 reference stations in Switzerland and northern Italy, obtaining a similar spatio-temporal correlation (0.79). However, several important differences limit the validity of this comparison: (i) Kim et al.'s study covers a much longer period (June 2018 to May 2020), including winter months, when satellite data is more frequently missing due to cloud cover—especially in complex alpine orography, which also affects the satellite’s ability to translate column densities into surface concentrations. (ii) Elevated regions like the Alps can introduce systematic biases in satellite-derived NO₂ due to vertical gradients in NO₂ distribution and reduced sensitivity near the surface.  (iii) Additionally, the amount of stations is much higher in the Kim et al. study (340 stations vs. 24 in Madrid), making their results spatially and statistically more robust. I suggest the authors reconsider the framing of the comparison or add more nuance to highlight the limitations and contextual differences that affect model performance in each case.

Minor comments:

1. pg 9 line 199: “See 0”
2. pg 17 line 360 “Sect. 0.”
3. pg 21 line 447: correct “Sect. 0”
4. pg 26, the manuscript states that "Direct assimilation of NO₂ satellite observations is not very useful due to the relatively short lifetime of NO₂ [...]" I would suggest that the issue may not lie in the inherent utility of the data, but rather in how the data is adapted and integrated into the model.

Dear authors,

I would like to thank you for a very interesting read! I have a couple of questions that I'd like you to reflect on, but overall I am very happy with the quality of the manuscript and the described research.

Eq.1

Looking back at Figure 2, the assumption that b can be assumed constant along the perimeter of the city seems a bit optimistic? There is a factor of six difference in the concentration along the border in the north and the south of the city in March 2019, which suggests that a westerly or easterly wind would cause a much higher flux across the border in the south than in the north.

And relatedly, you write (L179): “Other sectoral emission, e.g. from industry, will be accounted for indirectly in either an increased background field or in additional residential emissions.” Such industrial sites are likely not equally distributed along the border, further increasing inhomogeneities in the background border flux.

So my question is the following: Why not instead discretize the border along the l and z, and apply the same dispersion kernel that is used inside the city?

L204

Is it reasonable to assume that residential emissions are similar during weekdays and in the weekend?

L230-L234

Can you explain a bit about how AERMOD treats dispersion through street canyons? If only one dispersion kernel is calculated for each combination of wind speed and direction, stability and boundary layer height, the model cannot deal with variations in the built environment or even in roughness length (which I imagine can vary a lot from the sparsely populated northern area to downtown Madrid), correct? Do you expect this to result in large errors?

L269-L270

The dispersion model calculates concentrations of NO_x but rather than assimilating NO_x measurements, you assimilate assimilate NO₂ measurements. You get these from the XGBoost algorithm, which introduces non-linearity to the system. While you explicitly mention that you ignore the dependence on O₃ (L85-L86 of the SI), there is also the dependence on e.g. the temperature and the SEA (L264-L264). How do you reconcile this with the fact that a Kalman Filter assumes a linear measurement operator H?

Minor comments

L199: “See 0?”

L575: Mijling (2000) should be Mijling (2020)

Mult: Please also check the manunscript for many different occurrences of Sec. 0.