The complete framework of ECD-Net is illustrated in Figure 1 . It is designed based on the classic U-Net architecture, incorporating several enhancements to improve its performance. Specifically, ECD-Net retains the fundamental encoder-decoder structure of U-Net, while introducing additional modules to better capture spatial and contextual information. These modifications enable the network to achieve more precise feature extraction and segmentation results.

The encoder comprises five stages that progressively extract features while reducing spatial resolution and enriching semantic information. Stage 1 applies a 3 × 3 convolutional layer (Conv 3 × 3), followed by batch normalization (BN) and the ReLU activation function, to extract initial features while preserving the input’s resolution. Stage 2 introduces downsampling with a Conv 3 × 3 (stride = 2) to capture deeper feature representations, complemented by additional convolutional and normalization layers for feature refinement. Stages 3 and 4 incorporate MSDA Modules, enabling the model to effectively capture multi-scale cloud features and extract comprehensive global semantics. In Stage 5, a MHSA Modules is employed to model global dependencies, extracting high-level semantic features that serve as the foundation for the decoder.

The decoder mirrors the encoder’s structure but focuses on progressively recovering the spatial dimensions of the feature maps to generate the cloud mask. Starting from the output of the encoder’s final stage, Stage 5 of the decoder utilizes transposed convolution (ConvTrans + BN, K = 2, S = 2) for upsampling and incorporates a MHSA Modules to integrate global semantic information. Stages 4 and 3 continue upsampling through transposed convolution, leveraging MSDA Modules to enhance feature refinement. Skip connections are employed to fuse features from corresponding encoder stages, ensuring information completeness and preserving spatial details. Stage 2 further restores feature resolution closer to the original image size. Finally, Stage 1 applies a 1 × 1 convolutional layer to produce the cloud mask, which is normalized to a probability distribution in the range [0, 1] using a Sigmoid function.

The depth design of the attention module in both the decoder and encoder is the same [2, 2, 1].

This paper introduces the MSDA [23] module and the MHSA [24] [25] module to obtain multi-scale and global features. As represented in Figure 2(a) , the input first passes through Conditional Positional Encoding (CPE) [26] to incorporate positional information. This is followed by Layer Normalization (LayerNorm) to stabilize training. Then, the MSDA is applied to capture semantic information at different scales, with the output added back to the input via a residual connection to preserve the original information. Layer normalization is applied again, and an MLP is utilized to enhance the feature representation. In the end, another residual connection is applied to complete the module’s output. The MHSA module ( Figure 2(b) ) follows a similar architecture to the MSDA module, leveraging the MHSA mechanism to model long-range dependencies within the input features.

As shown in Figure 3 , the structure of MSDA changes such that the inputs Q, K, and V are generated through linear layers. Next, the channels are divided into n distinct attention heads., utilizing a SWDA (Sliding Window-based Sparse Attention) [23] mechanism to each head with varying dilation rates.

The operation of SWDA can be described as follows: given three matrices—Q, K and V as inputs, SWDA computes attention scores for each query vector $q_{ij}$ at position (i, j) by selecting sparse K and V vectors within a sliding window centered at that point. The sparsity is controlled by the dilation rate r.

$x_{ij}=\text{Attention}\left(q_{ij},K_r,V_r\right)=\text{Softmax}\left(\frac{q_{ij}{K}_r^{\text{T}}}{\sqrt{d_k}}\right)V_r$ (1)

This paper adopts a multi-supervision approach to train ECD-Net, focusing on improving the model’s performance. During training, the loss function uses BCE Loss [27] , which calculates the pixel-wise binary classification error between the predicted results and the ground truth labels, effectively optimizing the model parameters. To further enhance the training performance, we design a multi-supervision loss function ( ${ℒ}_{MS}$) based on BCE Loss. This function provides supervision at multiple-path, helping the model learn richer feature representations and improving the network’s ability to adapt to complex scenarios.

Specifically, our ${ℒ}_{MS}$ computes the loss at each stage and combines them through a weighted sum, allowing the model to integrate feedback from different layers and avoid overly relying on features from any specific layer. With this multi-level and multi-scale supervision mechanism, the model can better learn both fine-grained details and global semantics in the image, further improving the accuracy and robustness of cloud detection.

The specific definition of this ${ℒ}_{MS}$ is as follows:

${ℒ}_{MS}\left(Y,\hat{Y}\right)={\displaystyle \underset{n=1}{\overset{5}{\sum }}{ℒ}_{BCE}\left(Y,{\hat{Y}}_n\right)}$ (2)

where the ${\hat{Y}}_n$ represents the cloud mask generated at each stage.

The performance of ECD-Net was evaluated using the GF1-WHU dataset [28] and the SPARCS dataset [29] . The GF1-WHU dataset includes 86 satellite images for training and 22 for testing, covering diverse cloud conditions and land types. To simplify the cloud detection task, cloud shadows in the cloud mask images were treated as background. RGB image patches with a resolution of 384 × 384 pixels were generated through cropping, resulting in 2012 training patches and 516 testing patches. During training, the dataset was randomly split into training and validation subsets in an 8:2 ratio. The SPARCS dataset consists of 80 RS images with a resolution of 1000 × 1000 pixels. These images were manually divided into 64 training images and 16 testing images based on different land surface types. The images were further cropped into patches of 384 × 384 pixels, yielding 576 patches for the training set and 144 patches for the testing set. During training, the training data was randomly split into training and validation subsets in a 9:1 ratio.

In this study, the proposed ECD-Net was trained using the PyTorch framework [30] and optimized with the AdamW optimizer [31] . A cosine annealing scheduler with linear warm-up was employed to alter the learning rate. During training, the batch size was set to 24, with an initial learning rate of 4e-4 and a weight decay of 1e-3. The experiments ran for a total of 150 epochs. All experiments were conducted on a Windows 11 operating system and executed on an NVIDIA GeForce RTX 4090 GPU. This paper uses accuracy [14] , Jaccard Index (Jaccard) [32] and F1-Score [33] to evaluate the proposed model.

According to the quantitative comparison results presented in Table 1 and Figure 4 , ECD-Net excels in key evaluation metrics such as accuracy (97.37%), Jaccard index (87.09%), and F1 score (93.10%), demonstrating its high precision and superior segmentation performance in cloud detection tasks. Despite its higher computational cost (GFLOPs of 55.711), which requires more computational resources, its outstanding performance makes it the model of choice for tasks requiring high accuracy. In comparison, AMCD-Net, with a GFLOPs of 37.187, strikes a better balance between performance and computational efficiency. While it slightly trails ECD-Net in accuracy and F1 score, it still delivers excellent results. U-Net, as a traditional network architecture, has a higher computational complexity (GFLOPs of 89.955), but still performs well regarding accuracy (96.92%) and F1 score (91.92%), making it a widely used baseline model for cloud detection tasks. RS-Net, with a reduced computational load (GFLOPs of 38.183), performs exceptionally well in optimizing computational efficiency. Its accuracy (96.98%) and F1 score (91.93%) are similar to U-Net’s, making it a more efficient alternative.

However, ECD-Net’s inference speed (88.58 FPS) is significantly lower compared to other methods. This can be attributed to the computationally intensive self-attention mechanism and the multi-path supervision strategy employed in the model. While these techniques effectively enhance detection accuracy, they also increase the computational burden, resulting in slower inference speed.

This paper compares the visualization results of four models across three different scenes, as shown in Figure 5 . In the thin cloud regions over the water scene, all methods display inconsistent rates of missed and false detections. Among them, ECD-Net performs the most accurately, with clear boundaries between the cloud layers and the background, showing minimal missed and false detections. Although AMCD-Net is slightly inferior to ECD-Net, it still maintains high segmentation accuracy, with well-defined boundaries and fewer errors. In contrast, U-Net and RS-Net show larger areas of missed detections.

In the land scene, where clouds and background are distinctly different in color and texture, all methods perform well in cloud detection. However, RS-Net’s performance is somewhat lacking, with less precise segmentation of the clouds.

In the grass scene, U-Net and RS-Net show significant areas of missed detections, while AMCD-Net experiences only a few missed detections. ECD-Net delivers the best segmentation results in this scene, accurately identifying most of the cloud layers.

Overall, ECD-Net performs the best across all scenes, followed by AMCD-Net. U-Net and RS-Net show some limitations in both accuracy and detail handling.

As Table 2 shows that, consistent with its performance on the GF1-WHU dataset, ECD-Net achieves the best results in terms of accuracy (96.32%), Jaccard index (81.75%), and F1-Score (89.96%), fully demonstrating its exceptional performance in cloud detection tasks. However, despite having parameters and GFLOPS similar to other models, the inference speed of ECD-Net (33.70 FPS) is significantly lower.

Overall, while ECD-Net excels in accuracy and detection capability, its slower inference speed may limit its applicability in real-time or high-efficiency scenarios. Therefore, future work should focus on optimizing the model for lightweight design and faster inference to enhance its practicality further.

As shown in Figure 6 , all methods accurately detect the cloud regions in the first row, though ECD-Net and AMCD-Net exhibit a small number of false positives. In the second row, ECD-Net demonstrates the highest precision in capturing the boundaries and shapes of the cloud regions. In the third row, the edge detail detection of U-Net, RS-Net, and AMCD-Net appears somewhat coarse, while

ECD-Net more clearly restores the distribution details of the clouds. The fourth row illustrates a scene with sparse and isolated thin clouds, where ECD-Net accurately captures the thin cloud edges but shows false positives in the background of cloud-free areas. Overall, ECD-Net demonstrates superior performance in detecting edge details and thin clouds.

To evaluate the effectiveness of multi-supervision in the ECD-Net model, we compared the performance of using BCE loss to supervise the cloud mask generated by the final layer with that of using multi-path supervision. The data in Table 3 reveals that integrating multi-path supervision yielded a 0.18% higher F1-score and 0.28% improvement in the Jaccard index compared to training with BCE loss alone. This improvement suggests that multi-path supervision provides more comprehensive guidance during training, significantly enhancing the model’s performance in cloud detection tasks.