CloudMViT: Cloud Classification Using Ground-Based Remote Sensing Imagery and a Lightweight Hybrid Architecture

Xu, Wei; Wu, Ningning; Feng, Lin

doi:10.5194/egusphere-2026-1512

Preprints

https://doi.org/10.5194/egusphere-2026-1512

Preprints

11 May 2026

| 11 May 2026

CloudMViT: Cloud Classification Using Ground-Based Remote Sensing Imagery and a Lightweight Hybrid Architecture

Wei Xu, Ningning Wu, and Lin Feng

Abstract. Ground-based remote sensing cloud image data can be used to analyze regional cloud type variation trends, thereby predicting future water resource supply capacity. However, existing cloud classification methods based on ground-based remote sensing imagery often suffer from limited recognition accuracy due to insufficient fine-grained feature extraction, and their large model parameter counts hinder deployment on embedded terminals. To address these issues, this study proposes CloudMViT, a lightweight hybrid network architecture fusing a dual-pooling channel attention module and cross-scale self-attention, which enhances both local and global feature representation of cloud images while optimizing computational efficiency. Specifically, the model suppresses sky background interference and strengthens cloud edge features via the dual-pooling channel attention module that combines global average pooling (GAP) and global max pooling (GMP); captures cross-channel detailed features (e.g., cirrus fibril structures and stratocumulus shadows) using depthwise separable convolution and a decoupling mechanism; and further reduces model parameters by introducing CloudGhost cascade compression technology through linear feature redundancy elimination.Experiments on the World Meteorological Organization (WMO)-compliant HBMCD (10 standard cloud genera) and GCD (7 sky conditions) datasets demonstrate that CloudMViT achieves classification accuracies of 98.81 % and 95.13 %, respectively, significantly outperforming lightweight models such as MobileViT and EfficientNet. Ablation experiments validate the effectiveness of the dual-pooling channel attention module (improving accuracy by 5.31 %) and the CloudGhost module (increasing inference speed by 50 %). When deployed on the RK3588 embedded platform, the INT8-quantized CloudMViT enables real-time inference, maintaining an accuracy of 94.79 % with only 0.47 MB of memory occupation. The proposed cloud classification method and hardware acceleration strategy provide a feasible solution for the development of portable ground-based cloud observation and classification devices.

Received: 19 Mar 2026 – Discussion started: 11 May 2026

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Wei Xu, Ningning Wu, and Lin Feng

Status: final response (author comments only)

RC1:
'Comment on egusphere-2026-1512', Anonymous Referee #1, 08 Jun 2026
This manuscript presents an impressive work on ground-based cloud classification. It proposes CloudMViT, an innovative lightweight hybrid architecture that effectively integrates a dual-pooling channel attention module and cross-scale self-attention. The model achieves quite high classification accuracies on two high-quality datasets, including the fully WMO-compliant HBMCD and the GCD dataset which contains multiple WMO-standard cloud genera, and demonstrates excellent practical viability through successful deployment on the RK3588 embedded platform. To further strengthen the manuscript's impact and clarity, adding the introduction to cloud categories and the cloud image datasets, providing a deeper physical interpretation of the model's behavior, a detailed analysis of quantization-induced accuracy loss, and clarifying a few points regarding symbol/table consistency would be beneficial. I recommend publication after these minor enhancements.
Major Comments
The manuscript should add a section to introduce the cloud categories for classification and the cloud image datasets.

The manuscript discusses the overall accuracy drop of 4.02% after INT8 quantization but does not analyze which cloud categories suffer the most significant loss from a physical perspective. Specifically, the recall of cirrocumulus (Cc) drops dramatically from 97.39% to 87.06%, yet the physical reason for this degradation is not explained. The authors should provide a detailed analysis of why cirrocumulus are more sensitive to quantization errors, and how the quantization process destroys these physical texture features. A quantitative comparison of recall changes across all cloud categories is strongly recommended.

The manuscript mentions that the ECA-DP module suppresses background interference via GMP, but does not explain from a physical perspective why traditional CNNs (e.g., MobileNet) struggle with blurred cloud-sky boundaries. Furthermore, how does the self-attention mechanism in CloudMViT physically capture such gradual transitions? The authors should supplement the description of the physical characteristics of cloud edges and clarify how specific modulesrespond to these physical properties.

The manuscript states that "DW convolution can extract subtle edge differences between cirrus (Ci) and cirrocumulus (Cc)," but does not specify how the physical texture features of these clouds are represented in the model. The authors should add a discussion explaining how specific modules (e.g., the local receptive field of depthwise separable convolution) physically match the scale characteristics of these special textures.

Specific Comments
1.Many citation marker were missing in the literature; for example, line 55 Wang et al.; and in Section 2 Method, during the introduction of each method I couldn’t see any references
2.In Section 3.2, the manuscript states "total number of iterations was 100." However, deep learning training typically uses "epochs" rather than "iterations." The authors should clarify the intended meaning and unify the terminology to avoid confusion.
3. In Fig. 1, σ points to the output of the sigmoid activation function. In Eq. (1), η=σ(Vkγ), where σ already denotes the sigmoid function. However, Fig. 1 also shows a multiplication operation (⊗) after σ, which is not reflected in Eq. (1). The authors should check and correct this inconsistency.
4. The Conclusion states that the inference time is "only 0.006 s," but Table 10 lists an inference time of 30.14 s for CloudMViT (total time). Readers cannot easily derive the per-image inference time of 0.006 s. The authors should verify and correct the data, or explicitly state the calculation basis for the per-image inference time.
5. In Eq. (4) and Eq. (5), the parameters band rare only stated as "set to 1 and 2" in the text, but their physical meaning or derivation is not explained. Readers cannot understand why these specific values were chosen. The authors should clarify the physical significance or the basis for determining these parameters.
6. Fig. 6 labels step1, step2, and step3. Section 2.5 describes them as "Step 1: Extraction of local detailed features,""Step 2: Extraction of global features," and "Step 3: Enhancement of local detailed features." However, step1 and step3 have identical structures in the figure, and step2 is positioned in the middle. The text states that step3 "repeats step1," but the spatial arrangement in the figure conflicts with the semantic meaning of "repetition." The authors should revise either the figure or the text to ensure consistency.
7. The "Inference Time/s" column in Table 10 shows large values (e.g., 30.14 s for CloudMViT), but the manuscript does not specify that this is the total time for the entire test set . Readers may mistakenly interpret it as the per-image inference time, leading to misunderstanding of the model's speed. The authors should clearly label the column as "Total Inference Time on test set" in the table header or footnote, and also provide the per-image inference time in the main text.
Citation: https://doi.org/10.5194/egusphere-2026-1512-RC1
- AC1:
  'Reply on RC1', Wei Xu, 12 Aug 2026
  Dear Anonymous Referee #1,
  We sincerely thank you for your thorough and constructive review of our manuscript. Your positive feedback on our work is greatly appreciated. We have carefully considered all your comments and suggestions, and we have revised the manuscript accordingly. Below, we provide a detailed point-by-point response to your concerns.
  Major Comments
  The manuscript should add a section to introduce the cloud categories for classification and the cloud image datasets.
  
  Answer: We greatly appreciate your valuable suggestion. In response, we have specifically added an introduction to the WMO cloud classification standards, supplemented by the physical and morphological characteristics of different cloud genera, within the Introduction section. Meanwhile, in Section 3.1, we have explicitly distinguished between the "original image set" and the "augmented image set" (as shown in Tables 4 and 5), and we also provided the physical motivations for the data augmentation techniques (e.g., simulating varying illuminations, sensor noises, etc.) to avoid any confusion regarding the sample sizes.
  The World Meteorological Organization (WMO) classifies clouds into 10 basic genera (e.g., cumulus, cumulonimbus, stratocumulus, stratus) based on their shape and structure, and has published the latest International Cloud Atlas. As cloud density gradually decreases, the boundary between clouds and the sky becomes blurred, and clouds exhibit complex textural features—posing significant challenges for ground-based cloud image classification. These ten cloud genera exhibit distinct physical and morphological attributes, which inherently dictate varying classification difficulties. For instance, Cirrus and Cirrocumulus are characterized by delicate, high-frequency fibril structures or wave-like patterns requiring fine-grained spatial feature extraction. In contrast, Stratocumulus and Stratus possess predominantly uniform, layered structures with indistinct edges, where classification relies more on subtle cross-channel intensity variations, such as shadow depth. Furthermore, Cumulonimbus is prone to confusion with Cumulus due to their similar visual appearance at lower altitudes and the rapid vertical development of the former under unstable conditions. These distinctive structural properties impose different requirements on the network's feature extraction capability, specifically in addressing background interference, preserving global morphological contexts, and accurately capturing fine local textures.
  The manuscript discusses the overall accuracy drop of 4.02% after INT8 quantization but does not analyze which cloud categories suffer the most significant loss from a physical perspective. Specifically, the recall of cirrocumulus (Cc) drops dramatically from 97.39% to 87.06%, yet the physical reason for this degradation is not explained. The authors should provide a detailed analysis of why cirrocumulus are more sensitive to quantization errors, and how the quantization process destroys these physical texture features. A quantitative comparison of recall changes across all cloud categories is strongly recommended.
  
  Answer: We greatly appreciate the reviewer for raising this highly insightful physical perspective. In response, we have conducted a more detailed analysis of our experimental data and incorporated a comprehensive comparative analysis of the recall variations across all cloud categories before and after quantization into the revised manuscript.
  Detailed analysis reveals that the recall of Cirrocumulus (Cc) drastically declines from 97.39% to 87.06% (a drop of 10.33%), and Stratocumulus (Sc) decreases from 98.98% to 93.02%, whereas Stratus (St) and clear sky (No) experience a marginal decline of only approximately 3%. The physical reason for this discrepancy lies in the fact that Cc and Sc exhibit high-frequency, dense textural patterns (Cc features tightly packed scaly or rippled structures, while Sc involves subtle shadow gradients between large cloud blocks). The INT8 quantization process compresses floating-point values into merely 256 discrete integer bins, severely rounding off and smoothing out the minute grayscale gradients and fine edge details between these pixels. Conversely, St and No are characterized by low-frequency, smooth physical features (extensive uniform veil-like structures or solid backgrounds), where the information is concentrated in the low-frequency range, making them far less susceptible to the destructive effects of quantization rounding errors. This physical analysis not only clarifies the inherent limitations of quantization in edge deployment but also provides a theoretical foundation for future optimizations on embedded platforms, such as adopting mixed-precision strategies (e.g., retaining FP16 for critical layers) or quantization-aware training (QAT) to mitigate the accuracy loss for high-frequency texture cloud classes.
  The manuscript mentions that the ECA-DP module suppresses background interference via GMP, but does not explain from a physical perspective why traditional CNNs (e.g., MobileNet) struggle with blurred cloud-sky boundaries. Furthermore, how does the self-attention mechanism in CloudMViT physically capture such gradual transitions? The authors should supplement the description of the physical characteristics of cloud edges and clarify how specific modulesrespond to these physical properties.
  
  Answer: We greatly appreciate the reviewer's insightful focus on the physical properties of cloud imagery. In response, we have added detailed explanations in Sections 2.1 and 2.2 of the revised manuscript to elucidate the physical characteristics of the "blurred cloud-sky boundaries" and the corresponding response mechanisms of our proposed modules.
  For ECA-DP: The blurred cloud-sky boundary is often a region with low-intensity contrast, making it difficult for traditional CNNs (e.g., MobileNet) to distinguish boundaries from the background due to their limited local receptive fields and lack of global context. Physically, the Global Max Pooling (GMP) in ECA-DP captures the maximum luminance or gradient variation near the cloud boundary, effectively highlighting the sharpest structural edges regardless of the overall sky brightness. This mechanism physically suppresses the uniform and low-variance sky background, focusing the network's attention on the abrupt transitions at the cloud edges.
  For Self-Attention: The gradual transition from the sky to the cloud edge is not a hard line but a continuous gradient. While GMP identifies sharp edges, the self-attention mechanism physically captures this gradual transition by computing correlations between distinct local patches and their surrounding regions. It dynamically constructs a global spatial context, where the attention weights for pixels located at the fuzzy boundary are adaptively adjusted based on their similarity to both the high-contrast cloud features and the low-contrast sky background. This enables the model to physically encode the 'softness' of the boundary and the overall morphological distribution of the clouds, rather than treating pixels as independent features.
  The manuscript states that "DW convolution can extract subtle edge differences between cirrus (Ci) and cirrocumulus (Cc)," but does not specify how the physical texture features of these clouds are represented in the model. The authors should add a discussion explaining how specific modules (e.g., the local receptive field of depthwise separable convolution) physically match the scale characteristics of these special textures.
  
  Answer: To further elucidate the physical basis for selecting these specific modules for fine-grained cloud texture extraction, we analyzed the synergistic matching mechanism between the network components and the unique physical textural characteristics of different cloud genera, particularly Cirrus (Ci) and Cirrocumulus (Cc).
  Physical textural differences: From a meteorological perspective, Cirrus (Ci) is characterized by sparse, directionally-oriented filamentous/fibrous structures that are typically spatially discontinuous, whereas Cirrocumulus (Cc) exhibits dense, periodic scaly or rippled textures with relatively uniform cloud-block sizes.
  Physical matching of Depthwise Convolution (DW): The 3×3 depthwise convolution kernel possesses a limited local receptive field, making it ideally suited for capturing the inherent periodic local variations characteristic of Cirrocumulus's scaly patterns. This kernel scans individual scales and detects minute grayscale fluctuations between adjacent cloud blocks. Subsequently, the Pointwise Convolution (PW) aggregates these local texture primitives across the channel dimension to encode a complete morphological representation of the rippled texture.
  Physical matching of Self-Attention and Patch Partitioning: The self-attention module employs a 2×2 patch partitioning strategy. For Cirrus (Ci), this specific partition scale physically corresponds to the typical cross-sectional width of sparse fibril filaments. Each 2×2 patch can capture local, subtle fragments of the filaments. More importantly, the subsequent self-attention calculation computes long-range global dependencies among these scattered fragments. This mechanism effectively "reconstructs" the complete continuity of individual filaments across the full spatial extent of the image—a capability lacking in traditional CNNs due to their restricted receptive fields.
  Through this synergistic mechanism, the model achieves precise physical scale alignment at the cloud texture level, significantly enhancing the discriminative accuracy between the visually similar Ci and Cc categories.
  The above analysis has been incorporated into Section 2.3 "DW-PW Collaborative Feature Decoupling" of the revised manuscript.
  Specific Comments
  1.Many citation marker were missing in the literature; for example, line 55 Wang et al.; and in Section 2 Method, during the introduction of each method I couldn’t see any references
  Answer: We acknowledge that the foundational modules in the Methodology section lacked proper references. To address this issue, we have added the corresponding original citations for the core mechanisms. Specifically, we have cited [Vaswani et al., 2017] for the self-attention mechanism in Section 2.2, [Wang et al., 2020] for the ECA module in Section 2.1, and [Han et al., 2020] for the Ghost module in Section 2.4. The corresponding text has been revised accordingly.
  2.In Section 3.2, the manuscript states "total number of iterations was 100." However, deep learning training typically uses "epochs" rather than "iterations." The authors should clarify the intended meaning and unify the terminology to avoid confusion.
  Answer: We thank you for pointing out this terminology issue. In deep learning, the term "epochs" is conventionally used to denote the number of times the model iterates over the entire training dataset. To avoid confusion, we have revised the phrase "total number of iterations was 100" in Section 3.2 to "the total number of training epochs was set to 100," ensuring consistency with standard deep learning terminology.
  In Fig. 1, σ points to the output of the sigmoid activation function. In Eq. (1), η=σ(Vkγ), where σ already denotes the sigmoid function. However, Fig. 1 also shows a multiplication operation (⊗) after σ, which is not reflected in Eq. (1). The authors should check and correct this inconsistency.
  
  Answer: We sincerely thank you for pointing out the detailed issue of symbol inconsistency between Eq. (1) and Fig. 1. We fully agree with this comment and have revised the manuscript accordingly, with the following corrections made: The original phrase "these attention weights σ" in the main text was inaccurate, as σ represents the sigmoid activation function rather than the attention weights. We have corrected it to "these attention weights η" to ensure that the textual description is consistent with the symbols used in Eq. (1) and Fig. 1.
  The aforementioned revisions guarantee complete symbol uniformity and logical correspondence among Eq. (1), Fig. 1, and the main text, thereby avoiding any potential confusion for readers. Please refer to Section 2.1 and Eq. (1) in the revised manuscript for the specific changes.
  The Conclusion states that the inference time is "only 0.006 s," but Table 10 lists an inference time of 30.14 s for CloudMViT (total time). Readers cannot easily derive the per-image inference time of 0.006 s. The authors should verify and correct the data, or explicitly state the calculation basis for the per-image inference time.
  
  Answer: We sincerely thank you for identifying this inconsistency in the reported inference time. The reviewer's calculation is absolutely correct.
  Upon re-examination, the 30.14 s reported in Table 10 is indeed the total inference time on the entire PC-side test set, which contains 25,118 images. Therefore, the correct per-image inference time should be calculated as: The original 0.006 s stated in the Conclusion was erroneously derived using the test set size of the subsequent embedded deployment experiment (5,028 images), i.e., 30.14/5,028≈0.006. This cross-referencing of datasets indeed led to a misleading figure and confusion.
  We have made the following corrections in the revised manuscript: the GPU per‑image inference time stated in the Conclusion has been revised to 0.0012 s (1.2 ms), with a clear note that this value is calculated based on the full test set of 25,118 images. Meanwhile, in Table 10, the original heading “Inference Time” has been changed to “Total Inference Time on Test Set” to ensure full consistency between the table and the main text. We greatly appreciate the reviewer’s rigorous attention to detail, which has significantly improved the data consistency of our paper.
  In Eq. (4) and Eq. (5), the parameters band rare only stated as "set to 1 and 2" in the text, but their physical meaning or derivation is not explained. Readers cannot understand why these specific values were chosen. The authors should clarify the physical significance or the basis for determining these parameters.
  
  Answer: We sincerely thank you for pointing out the lack of explanation regarding the parameters in Eqs. (4) and (5). We fully agree that the derivation and basis for setting and were insufficient in the original manuscript.
  In the revised manuscript, we have added a clarifying paragraph in Section 2.1 as follows:
  （1）Definition of parameters: We explicitly stated that and are hyperparameters controlling the nonlinear mapping between the channel dimension and the 1D convolution kernel size . They are not physical constants specific to cloud physics.
  (2) Basis of selection: The mapping function and the specific values (r=2 and b=1) originate from the empirical optimization in the ECA-Net paper (Wang et al., CVPR 2020). According to the original study, these values were determined through grid search. The configuration r=2 and b=1 maps common channel dimensions (e.g., 16 to 1024) to appropriate odd kernel sizes (e.g., 3, 5, 7), effectively capturing local cross-channel interactions without introducing extra parameter redundancy.
  (3) Applicability: Since the ECA mechanism and this mapping strategy have been widely validated for general computer vision tasks, we directly adopted this baseline configuration in our ECA-DP module and verified its effectiveness for the specific task of cloud image classification.
  We have updated the text in the revised manuscript to provide the necessary context for readers, and we greatly appreciate this valuable suggestion.
  
  Fig. 6 labels step1, step2, and step3. Section 2.5 describes them as "Step 1: Extraction of local detailed features,""Step 2: Extraction of global features," and "Step 3: Enhancement of local detailed features." However, step1 and step3 have identical structures in the figure, and step2 is positioned in the middle. The text states that step3 "repeats step1," but the spatial arrangement in the figure conflicts with the semantic meaning of "repetition." The authors should revise either the figure or the text to ensure consistency.
  
  Answer: We sincerely thank you for pointing out the semantic inconsistency between Fig. 6 and the text regarding the term "repeated". We fully agree that the word "repeated" might mislead readers into thinking Step 3 is a cyclic or feedback mechanism, which contradicts the forward cascaded layout shown in the figure.
  In the revised manuscript, we have updated the description in Section 2.5 as follows:
  Avoided the term "repeat": We replaced "The same steps as Step 1 are repeated" with a more precise statement.
  Clarified the structural relationship: We now explicitly state that "Step 3 adopts the identical architectural design as Step 1". We also clarify that although the internal modules (PW-Conv, BatchNorm, SiLU, DW-Conv, BatchNorm, SiLU, and CloudGhost) are identical, Step 3 is positioned at a deeper stage of the network. Its purpose is to further refine the local detailed features before the final classification stage.
  This revision accurately reflects the forward, cascaded structure shown in Fig. 6 while eliminating the semantic ambiguity. No modifications to the figure itself are required. We have updated the text accordingly in the revised manuscript.
  The "Inference Time/s" column in Table 10 shows large values (e.g., 30.14 s for CloudMViT), but the manuscript does not specify that this is the total time for the entire test set . Readers may mistakenly interpret it as the per-image inference time, leading to misunderstanding of the model's speed. The authors should clearly label the column as "Total Inference Time on test set" in the table header or footnote, and also provide the per-image inference time in the main text.
  
  Answer: We thank you for pointing out the potential ambiguity in Table 10. We completely agree that the value 30.14 s listed in the table represents the total inference time on the entire test set (containing 25,118 images), not the per-image inference time. Without a clear label, this could indeed mislead readers.
  To resolve this issue, we have made the following corrections in the revised manuscript:
  Modified the table header: The column Inference Time/s has been renamed to Total Inference Time on Test Set (s).
  Updated the main text: We have revised the statements in the Conclusion to explicitly mention both the per‑image inference time and the total inference time,explicitly mention the per-image inference time alongside the total time, ensuring consistency between the table and the description.
  We believe these revisions will effectively eliminate any ambiguity and help readers accurately interpret both the total and per-image inference speeds. We sincerely appreciate the reviewer's meticulous attention to detail.
  
  Citation: https://doi.org/10.5194/egusphere-2026-1512-AC1
RC2:
'Comment on egusphere-2026-1512', Anonymous Referee #2, 05 Jul 2026

General Comments
This paper addresses the bottleneck of deploying ground-based cloud classification models on resource-constrained edge devices by proposing CloudMViT, a lightweight hybrid network architecture. The authors innovatively design a dual-pooling channel attention module (ECA-DP) and a cross-scale self-attention mechanism, which effectively enhance the extraction of cloud edge features and long-range textural characteristics. Meanwhile, the combination of depthwise separable convolution and CloudGhost cascade compression significantly reduces both parameter count and computational complexity.
The authors conduct thorough ablation and comparative experiments on two public benchmark datasets, HBMCD and GCD, and further validate the INT8-quantized model on the RK3588 embedded platform, achieving 21.58 FPS. Overall, the work is logically structured and experimentally well designed, and the achieved balance between accuracy and efficiency demonstrates considerable engineering value.
However, I note that the manuscript still has room for improvement in terms of clarity in design details, mathematical rigor, and adherence to standard terminology. The following comments are intended to help the authors further strengthen this otherwise promising contribution.
Major Issues ：
1、Inconsistent dataset sample numbers require clarification and justification of augmentation strategies.
The main text clearly states that the HBMCD dataset contains 25,118 images and the GCD dataset contains 12,000 images. However, in Tables 4 and 5, the "Total Samples" columns sum to 125,590 and 60,000, respectively. It is evident that the tables report post‑augmentation totals, whereas the text reports the original captured sample sizes. The authors should explicitly distinguish between the "original image set" and the "augmented image set" in both the text and the table captions. In addition, a brief justification should be provided for the four augmentation techniques (rotation, contrast enhancement, Gaussian noise, gamma correction).
2、The mathematical formulation of Global Max Pooling (GMP) is not rigorous.
The physical meaning of global max pooling is to take the maximum value over the entire spatial feature map in order to capture the most salient feature responses. However, Eq. (2) adopts a 'sum-then-max' formulation, which is effectively equivalent to taking the maximum of a scalar (the result of global average pooling) again. This completely defeats the original physical intention of GMP in highlighting edge information from locally strong activation regions. This equation should be carefully checked and corrected.
3、The assumption of a '1/2 proportion of original features' in the CloudGhost module lacks justification.
In Section 2.4, when deriving the compression ratio of the CloudGhost module, the authors directly assume in Eqs. (13) and (14) that 'the proportion of original features in the output cloud features is 1/2.' However, no rationale is provided for this specific choice. It is suggested that the authors clarify the basis for determining this ratio.
4、The introduction lacks a systematic summary of the core innovations.
The introduction effectively reviews the evolution of cloud classification methods and identifies three key challenges, but it ends abruptly without explicitly synthesizing how the proposed work addresses these challenges. I strongly recommend adding a dedicated paragraph at the end of the introduction that concisely summarises the main contributions. Such a summary would significantly improve the readability and narrative coherence of the paper.
Minor Issues:
1、The manuscript uses names such as exSwish, exLayerNorm, and exSoftmax (in Table 2). These should be clearly identified as standard operators (SiLU, LayerNorm, Softmax) at their first occurrence. I suggest using the format "SiLU (exSwish)" in the operator column of Table 2 to improve accessibility for a broader readership.
2、Throughout the paper (Tables 10 and 11), units such as "FLOPs/M" and "G" are used ambiguously. To conform to community standards, please adopt MFLOPs (mega floating‑point operations) or GFLOPs (giga floating‑point operations) consistently.
3、In Section 2.1, immediately after Eq. (5), the phrase "Here, |t| odd denotes the odd integer closest to t" contains a duplicated "odd" and is stylistically awkward.
4、Section 2.1 states: "Finally, an activation function k is used to calculate channel‑wise weights for the cloud feature map, and these attention weights σ are multiplied by the original input feature map." This is misleading: k is the kernel size of the 1D convolution, not an activation function. The activation function is the sigmoid function σ. The sentence should be rephrased to accurately reflect that the sigmoid function generates the normalized attention weights.
5、In Sections 2.4 and the abstract, the authors refer to the transformations used within the CloudGhost module to generate redundant features as 'linear operations.' However, these transformations typically involve depthwise convolution followed by batch normalization (BN) and activation functions, and are therefore not purely linear. It is suggested that this terminology be standardized to more accurately reflect the actual operations.

Citation: https://doi.org/10.5194/egusphere-2026-1512-RC2
- AC2: 'Reply on RC2', Wei Xu, 12 Aug 2026
  
  Dear Anonymous Referee #2,
  We sincerely thank you for the positive assessment of our work. We greatly appreciate your recognition of the engineering value and thorough experimental design of this study. We also highly value the insightful comments regarding clarity, mathematical rigor, and standard terminology. In the revised manuscript, we have meticulously addressed all these points, as detailed in our point-by-point responses below.
  Major Issues ：
  1、Inconsistent dataset sample numbers require clarification and justification of augmentation strategies.
  The main text clearly states that the HBMCD dataset contains 25,118 images and the GCD dataset contains 12,000 images. However, in Tables 4 and 5, the "Total Samples" columns sum to 125,590 and 60,000, respectively. It is evident that the tables report post‑augmentation totals, whereas the text reports the original captured sample sizes. The authors should explicitly distinguish between the "original image set" and the "augmented image set" in both the text and the table captions. In addition, a brief justification should be provided for the four augmentation techniques (rotation, contrast enhancement, Gaussian noise, gamma correction).
  Answer: We thank you for pointing out the inconsistency in dataset sample numbers. In the revised manuscript, we have explicitly distinguished between the "original image set" and the "augmented image set." In Section 3.1, we clarify that the original HBMCD and GCD datasets contain 25,118 and 12,000 images, respectively, and we further state that after data augmentation, the total numbers of training samples reached 125,590 and 60,000, respectively. To eliminate any confusion, we have also added the (Augmented) label to the headers of Tables 4 and 5. Additionally, we provided a physical justification for the four augmentation techniques: random rotation, contrast enhancement, Gaussian noise, and gamma correction are employed to simulate varying atmospheric conditions, changing solar illuminations, sensor noises, and camera exposure differences encountered in real-world deployments, thereby effectively mitigating overfitting and enhancing model robustness against complex environments.
  We employed four data augmentation techniques, including image rotation, contrast enhancement, Gaussian noise addition, and gamma correction. These augmentation methods are physically motivated to simulate varying atmospheric conditions, changing solar illuminations, sensor noises, and camera exposure differences encountered in real-world deployments, thereby effectively mitigating overfitting and enhancing model robustness against complex environments. Consequently, the total numbers of samples for the HBMCD and GCD datasets after augmentation reached 125,590 and 60,000, respectively.
  2、The mathematical formulation of Global Max Pooling (GMP) is not rigorous.
  The physical meaning of global max pooling is to take the maximum value over the entire spatial feature map in order to capture the most salient feature responses. However, Eq. (2) adopts a 'sum-then-max' formulation, which is effectively equivalent to taking the maximum of a scalar (the result of global average pooling) again. This completely defeats the original physical intention of GMP in highlighting edge information from locally strong activation regions. This equation should be carefully checked and corrected.
  Answer: We sincerely thank you for pointing out the lack of rigor in the mathematical formulation of GMP in Eq. (2). We carefully re-examined the equation, and you are absolutely correct that the previous 'sum-then-max' formulation entirely contradicts the physical intention of GMP, which is to take the maximum value over the spatial dimensions to extract salient features. We have fully accepted this suggestion and have corrected Eq. (2) to the standard spatial global max pooling formula in the revised manuscript: , where denote the spatial coordinates of the feature map, and represents the channel index.
  Specifically, we have made the following revisions in the updated manuscript:
  （1）Removed the summation symbol ∑ in Eq. (2) and directly applied the maximum operation over all spatial positions.
  （2）Added a physical explanation below the revised formula, clarifying that the operation retains the strongest activations and prevents edge information from being smoothed out by averaging.
  （3）Included the following clarification: "Here, denotes the maximum value across all spatial positions (i,j) in the feature map. This formulation physically aligns with the purpose of Global Max Pooling (GMP), which captures the most strongly activated features to highlight edge information, preventing critical details from being smoothed out by averaging."
  
  3、The assumption of a '1/2 proportion of original features' in the CloudGhost module lacks justification.
  In Section 2.4, when deriving the compression ratio of the CloudGhost module, the authors directly assume in Eqs. (13) and (14) that 'the proportion of original features in the output cloud features is 1/2.' However, no rationale is provided for this specific choice. It is suggested that the authors clarify the basis for determining this ratio.
  Answer: We thank you for pointing out that the assumption of a 1/2 proportion in the CloudGhost module lacked justification. In the revised manuscript, we have added a detailed explanation in Section 2.4. This ratio originates from the core design philosophy of the Ghost module (Han et al., CVPR 2020): generating half of the output feature maps via cheap operations and the other half via pointwise convolution achieves an optimal trade-off between parameter efficiency and representational capacity. Previous studies have also validated that this empirical split significantly reduces FLOPs while effectively preserving the feature diversity essential for fine-grained texture recognition. We have explicitly cited the original GhostNet literature to support this design choice.
  4、The introduction lacks a systematic summary of the core innovations.
  The introduction effectively reviews the evolution of cloud classification methods and identifies three key challenges, but it ends abruptly without explicitly synthesizing how the proposed work addresses these challenges. I strongly recommend adding a dedicated paragraph at the end of the introduction that concisely summarises the main contributions. Such a summary would significantly improve the readability and narrative coherence of the paper.
  Answer: We greatly appreciate your constructive suggestion. We fully agree with your observation that the original Introduction ended too abruptly after presenting the key challenges. In the revised manuscript, we have added a dedicated paragraph at the end of the Introduction to provide a systematic summary of the key challenges and core innovations, as follows:
  In summary, current ground-based remote sensing cloud classification algorithms mainly face two key challenges: (1) insufficient fine-grained feature extraction leading to low recall rates for certain cloud genera; (2) excessive model parameters hindering deployment on mobile meteorological observation devices. To systematically tackle these issues while fully harnessing the physical characteristics of various cloud types, the proposed CloudMViT framework is deliberately designed to align with these specific challenges. Specifically, the ECA-DP module employs a dual-pooling strategy to suppress uniform sky background while highlighting high-contrast edge features, effectively tackling the blurred cloud-sky boundaries seen in layered clouds and the fragmented, non-contiguous cloud blocks separated by clear-sky backgrounds. The cross-scale self-attention mechanism is incorporated to build long-range spatial dependencies beyond local receptive fields, thereby grasping the elongated, globally distributed fibril structures of Cirrus-like clouds. Meanwhile, the DW-PW collaborative decoupling module extracts precise local texture variations and inter-channel correlations, preventing feature loss from subtle cross-channel shadow variations in layered clouds. Finally, to meet the demands of portable meteorological observation for computational efficiency, the CloudGhost cascade compression strategy further reduces model parameters to ensure seamless deployment on edge devices. Through these targeted architectural innovations, this work aims to promote the intelligent development of ground-based remote sensing cloud observation methods and devices.
  Minor Issues:
  1、The manuscript uses names such as exSwish, exLayerNorm, and exSoftmax (in Table 2). These should be clearly identified as standard operators (SiLU, LayerNorm, Softmax) at their first occurrence. I suggest using the format "SiLU (exSwish)" in the operator column of Table 2 to improve accessibility for a broader readership.
  Answer: Thank you for this suggestion. In the revised manuscript, we have updated the operator names in Table 2 to the standard format: "SiLU (exSwish)", "LayerNorm (exLayerNorm)", and "Softmax (exSoftmax)". This approach retains the original operator specifics generated by the framework while ensuring easy identification of the standard operators for a broader readership.
  2、Throughout the paper (Tables 10 and 11), units such as "FLOPs/M" and "G" are used ambiguously. To conform to community standards, please adopt MFLOPs (mega floating‑point operations) or GFLOPs (giga floating‑point operations) consistently.
  Answer: We thank you for your meticulous attention to this detail. We have strictly standardized the units for FLOPs in the revised manuscript. Specifically, the units in Table 10 have been unified to MFLOPs (mega floating-point operations), while the units in Table 11 have been unified to GFLOPs (giga floating-point operations), ensuring consistency in units throughout the manuscript.
  
  3、In Section 2.1, immediately after Eq. (5), the phrase "Here, |t| odd denotes the odd integer closest to t" contains a duplicated "odd" and is stylistically awkward.
  Answer: We thank you for pointing out this awkward phrasing. In the revised manuscript, we have corrected this sentence in Section 2.1 to a clearer and more natural expression: " Here, represents the nearest odd integer to the value t, ensuring that the kernel size of the 1D convolution is an odd number."
  4、Section 2.1 states: "Finally, an activation function k is used to calculate channel‑wise weights for the cloud feature map, and these attention weights σ are multiplied by the original input feature map." This is misleading: k is the kernel size of the 1D convolution, not an activation function. The activation function is the sigmoid function σ. The sentence should be rephrased to accurately reflect that the sigmoid function generates the normalized attention weights.
  Answer: We sincerely thank you for this precise correction. We fully accept this comment. The original sentence misleadingly referred to the 1D convolution kernel size kk as an activation function. We have revised this in Section 2.1 to: " Finally, a sigmoid activation function σ is applied to generate channel-wise attention weights η for the cloud feature map. These weights η are then multiplied element-wise with the original input feature map to produce the final output cloud feature map.”
  5、In Sections 2.4 and the abstract, the authors refer to the transformations used within the CloudGhost module to generate redundant features as 'linear operations.' However, these transformations typically involve depthwise convolution followed by batch normalization (BN) and activation functions, and are therefore not purely linear. It is suggested that this terminology be standardized to more accurately reflect the actual operations.
  Answer: We thank you for pointing out this terminology issue. We fully agree with your comment that the transformations are not purely 'linear' due to the inclusion of Batch Normalization and activation functions. Therefore, as suggested, we have replaced the term 'linear operations' with 'cheap operations' (or ' efficient CloudGhost ') throughout the manuscript, particularly in Section 2.4.
  In Section 2.4, we also explicitly state that these cheap operations typically consist of a depthwise convolution, followed by Batch Normalization (BN) and an activation function. We emphasize that although these transformations incorporate non-linear components, their effective computational cost remains significantly lower than that of standard convolutions.
  
  We sincerely hope that our revisions fully address your concerns. We believe that the manuscript has been significantly strengthened in terms of clarity, mathematical rigor, and terminology accuracy. We look forward to your positive consideration.
  Sincerely,
  
  Wei Xu
  Dear Anonymous Referee #2,
  We sincerely thank you for the positive assessment of our work. We greatly appreciate your recognition of the engineering value and thorough experimental design of this study. We also highly value the insightful comments regarding clarity, mathematical rigor, and standard terminology. In the revised manuscript, we have meticulously addressed all these points, as detailed in our point-by-point responses below.
  Major Issues ：
  1、Inconsistent dataset sample numbers require clarification and justification of augmentation strategies.
  The main text clearly states that the HBMCD dataset contains 25,118 images and the GCD dataset contains 12,000 images. However, in Tables 4 and 5, the "Total Samples" columns sum to 125,590 and 60,000, respectively. It is evident that the tables report post‑augmentation totals, whereas the text reports the original captured sample sizes. The authors should explicitly distinguish between the "original image set" and the "augmented image set" in both the text and the table captions. In addition, a brief justification should be provided for the four augmentation techniques (rotation, contrast enhancement, Gaussian noise, gamma correction).
  Answer: We thank you for pointing out the inconsistency in dataset sample numbers. In the revised manuscript, we have explicitly distinguished between the "original image set" and the "augmented image set." In Section 3.1, we clarify that the original HBMCD and GCD datasets contain 25,118 and 12,000 images, respectively, and we further state that after data augmentation, the total numbers of training samples reached 125,590 and 60,000, respectively. To eliminate any confusion, we have also added the (Augmented) label to the headers of Tables 4 and 5. Additionally, we provided a physical justification for the four augmentation techniques: random rotation, contrast enhancement, Gaussian noise, and gamma correction are employed to simulate varying atmospheric conditions, changing solar illuminations, sensor noises, and camera exposure differences encountered in real-world deployments, thereby effectively mitigating overfitting and enhancing model robustness against complex environments.
  We employed four data augmentation techniques, including image rotation, contrast enhancement, Gaussian noise addition, and gamma correction. These augmentation methods are physically motivated to simulate varying atmospheric conditions, changing solar illuminations, sensor noises, and camera exposure differences encountered in real-world deployments, thereby effectively mitigating overfitting and enhancing model robustness against complex environments. Consequently, the total numbers of samples for the HBMCD and GCD datasets after augmentation reached 125,590 and 60,000, respectively.
  2、The mathematical formulation of Global Max Pooling (GMP) is not rigorous.
  The physical meaning of global max pooling is to take the maximum value over the entire spatial feature map in order to capture the most salient feature responses. However, Eq. (2) adopts a 'sum-then-max' formulation, which is effectively equivalent to taking the maximum of a scalar (the result of global average pooling) again. This completely defeats the original physical intention of GMP in highlighting edge information from locally strong activation regions. This equation should be carefully checked and corrected.
  Answer: We sincerely thank you for pointing out the lack of rigor in the mathematical formulation of GMP in Eq. (2). We carefully re-examined the equation, and you are absolutely correct that the previous 'sum-then-max' formulation entirely contradicts the physical intention of GMP, which is to take the maximum value over the spatial dimensions to extract salient features. We have fully accepted this suggestion and have corrected Eq. (2) to the standard spatial global max pooling formula in the revised manuscript: , where denote the spatial coordinates of the feature map, and represents the channel index.
  Specifically, we have made the following revisions in the updated manuscript:
  （1）Removed the summation symbol ∑ in Eq. (2) and directly applied the maximum operation over all spatial positions.
  （2）Added a physical explanation below the revised formula, clarifying that the operation retains the strongest activations and prevents edge information from being smoothed out by averaging.
  （3）Included the following clarification: "Here, denotes the maximum value across all spatial positions (i,j) in the feature map. This formulation physically aligns with the purpose of Global Max Pooling (GMP), which captures the most strongly activated features to highlight edge information, preventing critical details from being smoothed out by averaging."
  
  3、The assumption of a '1/2 proportion of original features' in the CloudGhost module lacks justification.
  In Section 2.4, when deriving the compression ratio of the CloudGhost module, the authors directly assume in Eqs. (13) and (14) that 'the proportion of original features in the output cloud features is 1/2.' However, no rationale is provided for this specific choice. It is suggested that the authors clarify the basis for determining this ratio.
  Answer: We thank you for pointing out that the assumption of a 1/2 proportion in the CloudGhost module lacked justification. In the revised manuscript, we have added a detailed explanation in Section 2.4. This ratio originates from the core design philosophy of the Ghost module (Han et al., CVPR 2020): generating half of the output feature maps via cheap operations and the other half via pointwise convolution achieves an optimal trade-off between parameter efficiency and representational capacity. Previous studies have also validated that this empirical split significantly reduces FLOPs while effectively preserving the feature diversity essential for fine-grained texture recognition. We have explicitly cited the original GhostNet literature to support this design choice.
  4、The introduction lacks a systematic summary of the core innovations.
  The introduction effectively reviews the evolution of cloud classification methods and identifies three key challenges, but it ends abruptly without explicitly synthesizing how the proposed work addresses these challenges. I strongly recommend adding a dedicated paragraph at the end of the introduction that concisely summarises the main contributions. Such a summary would significantly improve the readability and narrative coherence of the paper.
  Answer: We greatly appreciate your constructive suggestion. We fully agree with your observation that the original Introduction ended too abruptly after presenting the key challenges. In the revised manuscript, we have added a dedicated paragraph at the end of the Introduction to provide a systematic summary of the key challenges and core innovations, as follows:
  In summary, current ground-based remote sensing cloud classification algorithms mainly face two key challenges: (1) insufficient fine-grained feature extraction leading to low recall rates for certain cloud genera; (2) excessive model parameters hindering deployment on mobile meteorological observation devices. To systematically tackle these issues while fully harnessing the physical characteristics of various cloud types, the proposed CloudMViT framework is deliberately designed to align with these specific challenges. Specifically, the ECA-DP module employs a dual-pooling strategy to suppress uniform sky background while highlighting high-contrast edge features, effectively tackling the blurred cloud-sky boundaries seen in layered clouds and the fragmented, non-contiguous cloud blocks separated by clear-sky backgrounds. The cross-scale self-attention mechanism is incorporated to build long-range spatial dependencies beyond local receptive fields, thereby grasping the elongated, globally distributed fibril structures of Cirrus-like clouds. Meanwhile, the DW-PW collaborative decoupling module extracts precise local texture variations and inter-channel correlations, preventing feature loss from subtle cross-channel shadow variations in layered clouds. Finally, to meet the demands of portable meteorological observation for computational efficiency, the CloudGhost cascade compression strategy further reduces model parameters to ensure seamless deployment on edge devices. Through these targeted architectural innovations, this work aims to promote the intelligent development of ground-based remote sensing cloud observation methods and devices.
  Minor Issues:
  1、The manuscript uses names such as exSwish, exLayerNorm, and exSoftmax (in Table 2). These should be clearly identified as standard operators (SiLU, LayerNorm, Softmax) at their first occurrence. I suggest using the format "SiLU (exSwish)" in the operator column of Table 2 to improve accessibility for a broader readership.
  Answer: Thank you for this suggestion. In the revised manuscript, we have updated the operator names in Table 2 to the standard format: "SiLU (exSwish)", "LayerNorm (exLayerNorm)", and "Softmax (exSoftmax)". This approach retains the original operator specifics generated by the framework while ensuring easy identification of the standard operators for a broader readership.
  2、Throughout the paper (Tables 10 and 11), units such as "FLOPs/M" and "G" are used ambiguously. To conform to community standards, please adopt MFLOPs (mega floating‑point operations) or GFLOPs (giga floating‑point operations) consistently.
  Answer: We thank you for your meticulous attention to this detail. We have strictly standardized the units for FLOPs in the revised manuscript. Specifically, the units in Table 10 have been unified to MFLOPs (mega floating-point operations), while the units in Table 11 have been unified to GFLOPs (giga floating-point operations), ensuring consistency in units throughout the manuscript.
  
  3、In Section 2.1, immediately after Eq. (5), the phrase "Here, |t| odd denotes the odd integer closest to t" contains a duplicated "odd" and is stylistically awkward.
  Answer: We thank you for pointing out this awkward phrasing. In the revised manuscript, we have corrected this sentence in Section 2.1 to a clearer and more natural expression: " Here, represents the nearest odd integer to the value t, ensuring that the kernel size of the 1D convolution is an odd number."
  4、Section 2.1 states: "Finally, an activation function k is used to calculate channel‑wise weights for the cloud feature map, and these attention weights σ are multiplied by the original input feature map." This is misleading: k is the kernel size of the 1D convolution, not an activation function. The activation function is the sigmoid function σ. The sentence should be rephrased to accurately reflect that the sigmoid function generates the normalized attention weights.
  Answer: We sincerely thank you for this precise correction. We fully accept this comment. The original sentence misleadingly referred to the 1D convolution kernel size kk as an activation function. We have revised this in Section 2.1 to: " Finally, a sigmoid activation function σ is applied to generate channel-wise attention weights η for the cloud feature map. These weights η are then multiplied element-wise with the original input feature map to produce the final output cloud feature map.”
  5、In Sections 2.4 and the abstract, the authors refer to the transformations used within the CloudGhost module to generate redundant features as 'linear operations.' However, these transformations typically involve depthwise convolution followed by batch normalization (BN) and activation functions, and are therefore not purely linear. It is suggested that this terminology be standardized to more accurately reflect the actual operations.
  Answer: We thank you for pointing out this terminology issue. We fully agree with your comment that the transformations are not purely 'linear' due to the inclusion of Batch Normalization and activation functions. Therefore, as suggested, we have replaced the term 'linear operations' with 'cheap operations' (or ' efficient CloudGhost ') throughout the manuscript, particularly in Section 2.4.
  In Section 2.4, we also explicitly state that these cheap operations typically consist of a depthwise convolution, followed by Batch Normalization (BN) and an activation function. We emphasize that although these transformations incorporate non-linear components, their effective computational cost remains significantly lower than that of standard convolutions.
  
  We sincerely hope that our revisions fully address your concerns. We believe that the manuscript has been significantly strengthened in terms of clarity, mathematical rigor, and terminology accuracy. We look forward to your positive consideration.
  Sincerely,
  
  Wei Xu
  
  Citation: https://doi.org/10.5194/egusphere-2026-1512-AC2

Wei Xu, Ningning Wu, and Lin Feng

Viewed

Total article views: 325 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
259	48	18	325	26	23

HTML: 259
PDF: 48
XML: 18
Total: 325
BibTeX: 26
EndNote: 23

Views and downloads (calculated since 11 May 2026)

Month	HTML	PDF	XML	Total
May 2026	199	27	11	237
Jun 2026	26	7	4	37
Jul 2026	24	10	3	37
Aug 2026	10	4	0	14

Cumulative views and downloads (calculated since 11 May 2026)

Month	HTML	PDF	XML	Total
May 2026	199	27	11	237
Jun 2026	26	7	4	37
Jul 2026	24	10	3	37
Aug 2026	10	4	0	14

Viewed (geographical distribution)

Total article views: 300 (including HTML, PDF, and XML) Thereof 300 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 30 Aug 2026

Short summary

Clouds shape Earth’s climate and water supply. Classifying them from ground-based images helps track regional weather. Existing models are either inaccurate or too large for portable devices. We present a lightweight model, CloudMViT, using dual-pooling channel attention and cross-scale self-attention for cloud classification. Experiments show higher accuracy and real-time, low-memory performance on embedded hardware. This work supports portable cloud observation for climate monitoring.


Total:	0
HTML:	0
PDF:	0
XML:	0