Cloud4D: Estimating Cloud Properties at a High Spatial and Temporal Resolution

We thank the reviewer for their feedback and appreciate that they find our approach "groundbreaking" and the evaluation to be comprehensive. Please find below our response to your comments.

Weakness 1. While the experimental results are promising, the paper does not provide error bars or confidence intervals for the reported metrics

The main source of randomness in our experiments comes from the seed used for the weight initialization and dataset sampling during training. The model takes over 280 H100 hours to train, so running multiple full training runs is not feasible. To test the model variability, we have included results below on the mean and the 2-sigma error over five runs with random seeds trained for 9k steps (the full training run takes 90k steps). The results show that our method is robust to randomness in initialization and data sampling.

Weakness 2. the implementation details of the wind retrieval method could be elaborated further.

Additional implementation details of the wind retrieval method are given in Appendix B.3. We are happy to expand on this if there is anything missing. We will also release our code, which includes the full implementation details.

Weakness 3. While the combination of techniques presented in the paper is novel, some components of the framework, such as the use of vision foundation models (e.g., DINOv2) and sparse transformers, have been explored in other computer vision applications

We thank the reviewer for acknowledging the novelty of our proposed system. We agree that foundational models like DINOv2 and sparse transformers are established tools. Indeed, our motivation was precisely to leverage their representational power to solve the challenging task of estimating 3D cloud properties from 2D images, which has not been done before.

Q1. Could the authors elaborate on the challenges and potential strategies for extending Cloud4D to other cloud types (e.g., stratocumulus, cirrus) and multi-layer clouds?

As most cloud fields follow a vertically thin and horizontally wide structure, our method can be applied to other cloud types too. This could be accomplished by generating synthetic data for other cloud types.

While our cloud layer estimation model is currently designed for single cloud layers, it could be extended to multiple cloud layers. For example, instead of predicting 2.5D cloud features $\mathbb{R}^{N_x \times N_y \times 3}$, you could estimate additional channels to represent higher layers $\mathbb{R}^{N_x \times N_y \times 3L}$, where L would be the number of cloud layers to predict and would realistically not need to be more than two or three.

Q2. How might the framework handle occlusions and distinguish between different cloud layers?

Our current model is designed for single cloud layers. See our comment to Q1 above for suggestions on how this could be extended.

Q3. The authors should consider providing error bars or other statistical measures for the reported metrics

See answer to Weakness 1 above.

Q4. Explaining the factors contributing to the variability (e.g., train/test splits, initialization) would strengthen the statistical significance of the findings

Train/test splits. To ensure minimal leakage between training and test sets for the synthetic part of the evaluation, we do the split based on different meteorological environments used to configure the LES-based generation of synthetic cloud data.

For training, we use the following three meteorological conditions: ARM [1], Cabauw [2], and RICO [3].

For testing, we use BOMEX [4].

Initialization. We initialize our model weights using a Xavier uniform distribution.

Q5. How does the explicit modeling of camera geometry in Cloud4D contribute to its performance compared to the implicit learning approaches used?

Our camera setup consists of sparse views that are spaced far apart. As a result, the input images from the different views will greatly differ. The explicit modelling of camera geometry makes it easier to learn how to predict full 3D cloud properties from the sparse input images.

Q6. Assumptions made in the homography-guided 2D-to-3D transformation and the limitations of the sparse-voxel refinement approach

Find below a list of assumptions in our model:

(1) The homography assumes a pinhole camera model with no radial distortion and accurate camera poses. These are common assumptions for 3D vision tasks.

(2) The backprojection of DINOv2 features assumes the cloud layer being reconstructed is not occluded. This is generally not a problem for single cloud layers, as the cameras are upwards facing, such that occlusions are minimized. Additionally, while one camera might be partially occluded, the features from other unoccluded cameras will help average out errors.

(3) The sparse transformer only refines in local areas of the estimated cloud from the cloud layer estimation model.

Q7. The authors could include experiments or simulations to validate the framework's performance under various environmental conditions

We have investigated our model's performance across different conditions in our real-world dataset, including different numbers of camera views, cloud coverages, cloud base heights, and cloud geometric thicknesses. These results are shown below, where the radar GT has been used to classify the cloud coverage, mean cloud base height, and mean cloud geometric thickness, over one-hour segments in our real-world dataset.

From these results, we highlight that the model’s performance slightly worsens for higher cloud base heights and also geometrically thicker clouds, but not to a significant extent.

Q8. How does Cloud4D handle partially obscured views or low visibility conditions?

Partially obscured views. To evaluate how Cloud4D might be affected by partially obscured views, we have included results below on how Cloud4D is affected by uniformly dropping random camera pairs during evaluation. The results show that predictions for cloud base heights and cloud top heights worsen as views are dropped, but that liquid water content and liquid water path predictions are less affected. The decreasing CBH and CTH accuracy with missing views matches established results in 3D vision, where increasing the number of views improves accuracy in 3D shape estimation.

Low visibility conditions. Our current implementation focuses on cases where no rain or fog is obscuring the views. Mitigating the impact of these conditions is left to future work (see Q9 for suggestions).

Q9. Are there any preprocessing steps or modifications to the framework that can mitigate the impact of these conditions?

Yes, here are some suggestions:

(1) The model could directly learn to be robust to low-visibility conditions by including such cases in the training data. The simulation of rain or fog could be implemented using existing techniques in Blender.

(2) The impact of low visibility conditions could be mitigated through preprocessing by using weather data to filter out data with rain or fog.

[1] Large-eddy simulation of the diurnal cycle of shallow cumulus convection over land, Brown et al. (Royal Meteorological Society 2002)

[2] An efficient parameterization for surface shortwave 3d radiative effects in large-eddy simulations of shallow cumulus clouds, Tijhuis et al. (Advances in Modeling Earth Systems 2023)

[3] Controls on precipitation and cloudiness in simulations of trade-wind cumulus as observed during rico, vanZanten et al. (Advances in Modeling Earth Systems 2011)

[4] A large eddy simulation intercomparison study of shallow cumulus convection, Siebesma et al. (Atmospheric Sciences 2003)

We thank the reviewer for their feedback and are happy to hear that they find the cloud estimation task very interesting and that they were impressed with the results in our satellite-based comparison. Please find below our response to your comments.

Dataset Generation

Q1. How was the data split between training and test sets?

For training, we use high-resolution LES, which is a state-of-the-art simulation approach for clouds. To ensure that our model generalizes across different conditions, we simulate a number of cases with different meteorological environments. In practice, we use a set of well-studied meteorological conditions and refer to these as “scenarios”.

To ensure minimal leakage between training and test sets for the synthetic part of the evaluation, we do the split based on the scenarios.

For training, we use the three scenarios mentioned in Line 180 of the paper: ARM [1], Cabauw [2], and RICO [3].

For testing, we use: BOMEX [4]

Q2. Please clarify the "three scenarios" mentioned in Line 180.

See answer to Q1 above.

Q3. Were different lighting conditions (such as sun azimuth) considered, and was lighting augmentation performed?

Yes, during the data generation, we randomly sample the sun position. We additionally sample the air density and the dust density, both of which also affect the lighting conditions.

Image augmentations are applied during training, where we vary the saturation, hue, and brightness. For full details, we refer to the training details that can be found in Appendix A of the supplemental material.

Q4. How accurately does the LES simulation reflect real-world conditions, and is this a standard methodology?

Yes, LES is a standard methodology for the simulation of high-resolution clouds. It is commonly used for the study of cloud development, and the simulations typically agree well with real-world measurements for cases with shallow convection (as in our setting) [4].

Q5. What types of real-world clouds cannot be simulated using cloud simulation?

There are no specific limits to the kinds of clouds that can be simulated. However, as the microphysics (behaviour of droplets and ice crystals) cannot be simulated directly, clouds involving complex precipitation processes or a mixture of liquid and ice are more challenging to simulate. We focus on non-precipitating shallow cumulus, which are commonly studied through simulation.

Q6. It would be beneficial to include visualizations comparing synthetic and real-world data.

Thank you for the suggestion. We will add an example of this to the supplementary.

Training and Fine-tuning

Q7. Has consideration been given to using Sentinel-2 imagery for model fine-tuning? Given the two-month real-world dataset and Sentinel-2's average 5-day revisiting time, it would be valuable to explore whether real-world fine-tuning improves performance.

Even with a 5-day revisiting time, Sentinel-2 only visits the area ~12 times over the two months our cameras were deployed. Of these visits, some are completely clear skies, and others are dominated by non-cumulus clouds. Unfortunately, this means that there is not enough data for fine-tuning at the current stage.

Q8. Does scaling up the training dataset size help?

Yes, from initial experiments, we found that increasing the number of scenarios used for training helped with the robustness of our model. This is likely due to the increased diversity in meteorological conditions in our training data resulting in better generalization.

Quantitative Evaluation

Q9. The paper lacks comprehensive quantitative evaluation of real-world data performance.

Table 1 in the paper shows quantitative metrics compared to radar retrievals using our two-month real-world dataset, where we have evaluated the performance on both retrieving physical cloud properties (liquid water content and liquid water path) and geometric properties (cloud occupancy, cloud base heights, cloud top heights).

As part of the rebuttal, we have additionally quantitatively investigated our model's error across different conditions in our real-world dataset, including different numbers of camera views, cloud coverages, cloud base heights, and cloud geometric thicknesses.

Our key takeaways from this investigation are the following:

(1) The CBH and CTH error increases as the number of views decreases, but the LWC and the LWP are less affected. The decreasing CBH and CTH accuracy with fewer views matches established results in 3D vision, where increasing the number of views improves accuracy in 3D shape estimation.

(2) We observed slightly worse performance for higher cloud base heights and also geometrically thicker clouds, but not to a significant extent.

For additional details, we have included the full numerical results in our response to reviewer XcpS.

Alternative Approaches

Q10. Have alternative methods such as NERF or DUST3R been considered for 3D estimation? How do these methods compare to the proposed approach?

DUSt3R only reconstructs surfaces, so it is not able to estimate cloud properties across the 3D volume. However, we have tried volumetric methods including MVSNeRF [5], IBRNet [6], and Gaussian Splatting [7]. We found that all of these models severely failed due to the limited number of camera views and because the domain is very different from that on which they were trained.

For MVSNeRF and IBRNet, we additionally tried fine-tuning on our synthetic data but found that this did not significantly improve results. This is not unexpected as these models are generally trained on synthetic datasets with much more than six camera views.

[1] Large-eddy simulation of the diurnal cycle of shallow cumulus convection over land, Brown et al. (Quarterly Journal of the Royal Meteorological Society 2002)

[2] An efficient parameterization for surface shortwave 3d radiative effects in large-eddy simulations of shallow cumulus clouds, Tijhuis et al. (Journal of Advances in Modeling Earth Systems 2023)

[3] Controls on precipitation and cloudiness in simulations of trade-wind cumulus as observed during rico, vanZanten et al. (Journal of Advances in Modeling Earth Systems 2011)

[4] A large eddy simulation intercomparison study of shallow cumulus convection, Siebesma et al. (Journal of the Atmospheric Sciences 2003)

[5] Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo, Chen et al. (ICCV 2021)

[6] IBRNet: Learning Multi-View Image-Based Rendering, Wang et al. (CVPR 2021)

[7] 3D Gaussian Splatting for Real-Time Radiance Field Rendering, Bernhard et al. (SIGGRAPH 2023)

We thank the reviewer for their feedback and appreciate that they acknowledge the difficulty of the target task and that they find our dataset to be a valuable contribution. Please find below our response to your comments.

Q1. The comparison with the VIP-CT method could have been described and explained in more detail

VIP-CT takes as input satellite-based images ($\mathbb{R}^{1 \times H \times W}$) and outputs a 3D field ($\mathbb{R}^{N_x \times N_y \times N_z}$). This is the same as our setup, with the only difference being that satellite imagery is grayscale while ours is RGB. This only requires a simple change in the number of channels for the first convolutional layer.

We use the publicly available training code for VIP-CT and instead train using our synthetic dataset, which consists of ground-based images rather than satellite-based images.

We will add this description to the paper. Thank you for pointing this out.

Q2. VIP-CT is designed for satellite images, how effective is this as a comparison model when applied to ground based camera systems?

A key difference between satellite images and ground-based images is that the viewing angles greatly differ across ground-based views. The lack of explicit camera geometry modeling in VIP-CT can therefore limit its effectiveness with our ground-based images. However, as there is no prior work that predicts volumetric 3D cloud properties from ground-based cameras, we use VIP-CT as the closest comparable model.

Q3. Are these just radar images does the model require optical images too?

To clarify, our model estimates full 3D physical cloud properties from only optical (RGB) images. No radar images are required, as we only use radar for the evaluation of our model.

Q4. If the model requires optical images, does this mean the system is only capable of operating during daytime?

The cameras we have deployed only operate during the daytime. However, we argue this is not a significant limitation as:

(1) Cumulus clouds generally require solar heat to form over land and are therefore mostly present during daytime.

(2) There are alternative cameras that can operate at night and are capable of observing clouds.

Q5. How scalable this method is if it relies on ground camera systems? I assume that means it cannot realistically have global coverage?

Our setup is scalable in the sense that it uses commodity cameras, which are cheap, so that new sites can be easily deployed. However, we do not intend for the setup to be deployed in a manner such that global areas are densely monitored by ground-based cameras. We envision them being deployed sparsely at sites of interest to cover large areas. This is because satellites are still the best option for global coverage, and the measurements from Cloud4D are complementary, providing high spatial and temporal resolution measurements for focused areas.

Q6. Benefits of investing efforts in furthering ground based systems as opposed to improving satellite-derived monitoring?

We see the ground-based system as complementary to satellites for two reasons. Firstly, the ground-based system has a much higher temporal and spatial resolution. Even with advanced satellites, it would be impossible to get data at 5s timesteps at 25m resolution, useful for fast forming convective clouds like cumulus. Secondly, in the visible spectrum, satellites observe mainly the tops of the clouds, while ground-based cameras are able to provide complementary observations of cloud bases.

Q7. Is there a middle ground where select ground based systems are used to generate training / evaluation data to refine satellite-based models?

Absolutely, we believe this is a key potential application of our method. Ground-based cameras can provide additional spatial detail and temporal context, while satellites have the advantage of global spatial coverage.

Weakness 1. However, the localised nature of this dataset left questions as to how generalisable this approach is.

We agree that testing the generalization to other sites would be useful. However, we would like to point out that our current model is trained exclusively on synthetic data and is never trained on data from our real-world test site. Only the camera poses in our synthetic data match our real-world setup; the other characteristics are decoupled from the real test site. Our synthetic dataset includes a wide variety of simulated cloud morphologies, atmospheric conditions, and sun positions that are not unique to our physical location. Therefore, the model is forced to learn the underlying principles of 3D cloud reconstruction from multi-view imagery, rather than overfitting to the specific visual features (e.g., background, typical weather patterns) of a single site. Our real-world site, therefore, acts as a generalization test in itself.

Weakness 2. From what I could see in the paper or supplementary material, the geographic location or season of this dataset was not presented.

We have not included the geographic location to minimize identifying information in the double-blind review process. However, we will include precise coordinates and timestamps together with the public release of our real-world dataset.

Weakness 3. It would be good to comment on how much variability there was in model error and the characteristics of the error

For the rebuttal, we have investigated our model's error across different conditions in our real-world dataset, including different numbers of camera views, cloud coverages, cloud base heights, and cloud geometric thicknesses.

Our key takeaways from this investigation are the following:

(1) The estimation of CBH and CTH worsens as the number of views decreases, but the LWC and the LWP are less affected. The decreasing CBH and CTH accuracy with fewer views matches established results in 3D vision, where increasing the number of views improves accuracy in 3D shape estimation.

(2) We observed slightly worse performance for higher cloud base heights and also geometrically thicker clouds, but not to a significant extent.

Please see our response to reviewer XcpS for the full numerical results.

We thank the reviewer for their thorough response. Please find below our response to your questions.

While I recognise the logistics and costs associated in setting up ground based camera systems, are there camera datasets which could be used as additional validation data (either public or private)?

To the best of our knowledge, SGP-ARM [1] is the only other camera dataset suitable as additional validation data. Below we show results validating our model and VIP-CT on a one-hour segment of cumulus clouds captured in the SGP-ARM dataset:

Our model shows similar accuracy on this dataset compared to the results on our real-world dataset. We note that VIP-CT performs significantly worse on this dataset, which can be attributed to the fact that it implicitly learns camera geometry and does not generalize to unseen camera poses. On the other hand, our method has no problem transferring to the SGP-ARM camera poses. The generalization across camera setups is due to the homography projection, which helps decouple the cloud estimation task from the camera poses as the learned module operates in world-space. This shows our model’s capability of generalizing to other sites and camera setups.

Do the authors have a strategy for scaling up the validation of the model?

The main effort in scaling up the validation would be capturing new data with suitable ground truth. We therefore believe that scaling up the validation is best done through two approaches:

(1) Single-site deployment over time: by deploying the cameras for longer periods, the validation scale will increase in both data quantity and cloud diversity (from seasonal weather changes).

(2) Deployment across multiple sites: this would increase the diversity of observed clouds through geographic changes.

What do the MSE values for the key outcome variables mean in terms of using model outputs for downstream analyses? Are they fit for purpose?

To put the error metrics into perspective, typical measurements taken with aircraft probes (which are the standard for in-situ measurements at multiple points throughout a cloud) have an uncertainty of around 10% - 15%. In comparison, our LWC relative error is around a similar range:

$\dfrac{\text{MAE(LWC of clouds)}}{\text{mean(LWC of clouds from radar GT)}} = \dfrac{0.029 gm^{-3}}{0.321 gm^{-3}} = 8.9\\%$

We would also like to highlight that there are no existing methods able to estimate dense 3D cloud fields at the spatial and temporal resolution of our method. Our method, therefore, fills an observational gap that is scientifically useful for downstream analyses.

I was also curious about Fig 3a - it seems that the model misses some spikes in reference 1D liquid water content. Have the authors explored why?

We believe that the main problem for those spikes is due to there being two cloud layers, where most clouds are at ~1000m, but with a small fraction at ~1750m. As discussed as part of our limitations (Section 7), our model is only trained for single cloud layers and thus misses the second cloud layer.

[1] Observing Clouds in 4D with Multiview Stereophotogrammetry, David M. Romps and Ruşen Öktem (BAMS 2018)

We thank the reviewer for their feedback, and we appreciate that they find our approach to be innovative and with significant potential. Please find below our response to your comments.

Q1. Important: Justification of 10% Relative Error

Thank you for pointing this out. We agree that it is important for this to be clear. The relative error specifically refers to our LWC predictions when compared to radar values and is calculated as

$\dfrac{\text{MAE(LWC of clouds)}}{\text{mean(LWC of clouds from radar GT)}} = \dfrac{0.029 gm^{-3}}{0.321 gm^{-3}} = 8.9\\%$

where the mean absolute error of our LWC predictions is the same as the value from Table 1 in the paper.

We will make this clearer in the paper.

Q2. Moderate: Isn’t there a need to address the domain gap between synthetic and real-world observations when deploying the model on real camera and radar data?

It would be ideal to train our model with real-world data, but there are currently no observing systems that are able to provide the full 3D observations of cloud fields at the spatial scale required. To fill this gap, we use high-resolution LES, which is a state-of-the-art simulation approach for clouds. To ensure that our model generalizes across different conditions, we simulate a number of cases with different meteorological environments.

Our results validate the effectiveness of LES-based training, showing that real-world data is not required, although it may improve results further. For future work, it could be interesting to investigate how real-world data might be integrated with simulation-based training.

Q3. Moderate: Did you consider implementing and comparing with other models (e.g., multi-view stereo, which is also referenced in your work)?

Multi-view stereo only reconstructs surfaces, so it is not able to estimate cloud properties across the 3D volume. However, we have tried volumetric methods including MVSNeRF [1], IBRNet [2], and Gaussian Splatting [3]. We found that all of these models severely failed due to the limited number of camera views and because the domain is very different from that on which they were trained.

For MVSNeRF and IBRNet, we additionally tried fine-tuning on our synthetic data but found that this did not significantly improve results. This is not unexpected as these models are generally trained on synthetic datasets with much more than six camera views.

Q4. Minor: In Section 3.1 and related equations, does the symbol "K" refer to the number of height planes sampled, the number of camera intrinsics, or both?

Thank you for pointing this out. K refers to the number of height planes sampled, and we will update the notation to avoid confusion with the camera intrinsics that are denoted with a bolded K. We denote the number of images (which are the same as the number of intrinsics) as N.

Weakness 1. Generalizability to other sites, camera setups, number of views or cloud regimes is not tested or discussed in depth. Of course, this is an inherent limitation of the method, since deploying many sites and setups requires significant effort.

We agree that testing the generalization to other sites would be useful. However, we would like to point out that our current model is trained exclusively on synthetic data and is never trained on data from our real-world test site. Only the camera poses in our synthetic data match our real-world setup; the other characteristics are decoupled from the real test site. Our synthetic dataset includes a wide variety of simulated cloud morphologies, atmospheric conditions, and sun positions that are not unique to our physical location. Therefore, the model is forced to learn the underlying principles of 3D cloud reconstruction from multi-view imagery, rather than overfitting to the specific visual features (e.g., background, typical weather patterns) of a single site. Our real-world site, therefore, acts as a generalization test in itself.

We have also investigated our model's performance across different conditions in our real-world dataset, including different numbers of camera views, cloud coverages, cloud base heights, and cloud geometric thicknesses. These results are included below, where camera pairs are randomly and uniformly dropped, and the radar GT has been used to classify the cloud coverage, mean cloud base height, and mean cloud geometric thickness, over one-hour segments in our real-world dataset.

Our key takeaways from this investigation are the following:

(1) The estimation of CBH and CTH worsens as the number of views decreases, but the LWC and the LWP are less affected. The decreasing CBH and CTH accuracy with dropped views matches established results in 3D vision, where increasing the number of views improves accuracy in 3D shape estimation.

(2) We observed slightly worse performance for higher cloud base heights and also geometrically thicker clouds, but not to a significant extent.

Weakness 2. The method relies on radar retrievals for ground truth, which themselves have limitations and may not always be available or accurate, especially for shallow clouds.

We agree that radars have their own biases and limitations. We would like to note that the liquid water content retrieval of the radar also includes information from a coincident lidar and is also scaled to match a microwave radiometer [4]. It is the best possible evaluation dataset for this deployment, but we note the possibility of errors in this data.

[1] Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo, Chen et al. (ICCV 2021)

[2] IBRNet: Learning Multi-View Image-Based Rendering, Wang et al. (CVPR 2021)

[3] 3D Gaussian Splatting for Real-Time Radiance Field Rendering, Bernhard et al. (SIGGRAPH 2023)

[4] Cloudnet: Continuous Evaluation of Cloud Profiles in Seven Operational Models Using Ground-Based Observations, Illingworth et al. (AMS 2007)

We thank the reviewer for their feedback and appreciate that they find our approach interesting and the target task to be very relevant and important. Please find below our response to your comments.

Q1. Could the authors provide more context about how the system is meant to be deployed?

We envision the following possible deployment scenarios:

(1) Small-scale at a single site: this would allow studying cloud physical properties and processes, such as formation and dynamics, which is not possible with existing instruments. It could allow the development and evaluation of parameterisations for weather and climate models. It could also provide a satellite validation site.

(2) Large-scale with sparse coverage: this would be deployed similarly to existing radar and meteorological instrument sites which span a large area but have non-overlapping coverage. The output of this deployment could feed into NWP models or be used for scientific studies.

(3) Large-scale with dense coverage: while dense coverage is more difficult, here we would aim to cover one weather model grid box (about 25km x 25km) to provide a 3D reconstruction of a portion of the cloud field spanning the full grid-box element. This would be useful for validating and improving weather/climate models.

Q2. If the deployment intention is the former, I would like to know about the requirements in terms of number of cameras required to cover, say, a 100kmx100km area, and which communication/collection infrastructure would be required

A dense deployment covering a 25km x 25km area would require the following:

Number of cameras. The most basic scaling of our deployment would be to duplicate our current camera setup across a grid. As our current setup is covering 5km x 5km using 6 cameras, scaling up to the sizes of a weather model gridbox ~(25km x 25km) would require 6 * 25 = 150 cameras. Given that our cameras are commodity cameras, this would still be cheaper than existing measurement devices such as radars. For future work, it could be of interest to investigate alternative camera setups for larger areas.

Infrastructure. The cameras have low power consumption and are powered by an external portable solar panel. Time synchronization and transfer of camera data to an external server is done using a standard 4G cellular connection through a Raspberry Pi.

Q3. I don’t see how the proposed method could be generalized to estimate cloud properties of different cloud types, namely thicker clouds higher in the atmosphere

We would like to highlight that while an individual cloud might be more vertically thick for other cloud types, the vertical thin and horizontally wide structure still holds for any overall cloud field.

This vertical structure is a property of the atmosphere in general and also motivates why weather models such as the Integrated Forecasting System (ECMWF) and GraphCast [1] have gridboxes which are also vertically thin and horizontally wide.

Q4. There are more suited satellite images available than Sentinel-2, e.g. PlanetScope from Planet Labs

PlanetScope is a good candidate for testing in the future, but as data access is only granted on request, we have chosen to compare against MODIS and Sentinel-2, as their data is publicly available and commonly used for cloud physics research.

[1] Learning skillful medium-range global weather forecasting, Lam et al. (Science 2023)