CloudBreaker: Breaking the Cloud Covers of Sentinel-2 Images using Multi-Stage Trained Conditional Flow Matching on Sentinel-1 Saleh Sakib Ahmed, Sara Nowreen, M. Sohel Rahman Abstract Cloud cover and nighttime conditions remain significant limitations in satellite-based remote sensing, often restricting the availability and usability of multi-spectral imagery. In contrast, Sentinel-1 radar images are unaffected by cloud cover and can provide consistent data regardless of weather or lighting conditions. To address the challenges of limited satellite imagery, we propose CloudBreaker, a novel framework that generates high-quality multi-spectral Sentinel-2 signals from Sentinel-1 data. This includes the reconstruction of optical (RGB) images as well as critical vegetation and water indices such as NDVI and NDWI.We employed a novel multi-stage training approach based on conditional latent flow matching and, to the best of our knowledge, are the first to integrate cosine scheduling with flow matching. CloudBreaker demonstrates strong performance, achieving a Fréchet Inception Distance (FID) score of 0.7432, indicating high fidelity and realism in the generated optical imagery. The model also achieved Structural Similarity Index Measure (SSIM) of 0.6156 for NDWI and 0.6874 for NDVI, indicating a high degree of structural similarity. This establishes CloudBreaker as a promising solution for a wide range of remote sensing applications where multi-spectral data is typically unavailable or unreliable. keywords Sentinel-1, Sentinel-2, NDWI, NDVI, Flow matching, scheduling, Generative AI, image-to-image translation 1 Introduction Cloud cover poses significant challenges in satellite-based Earth observation. Persistent cloudiness hampers the acquisition of clear satellite images, affecting various applications. Sentinel-2, part of the European Space Agency’s Copernicus Programme Copernicus Open Access Hub 2024 , has been providing high-resolution multispectral imagery, which includes additional signals, such as near-infrared (NIR) in addition to standard Optical (RGB) images. Using these additional signals, we can derive various important indices, such as the Normalized Difference Water Index (NDWI) and the Normalized Difference Vegetation Index (NDVI). These signals have applications in agriculture, forestry, water quality monitoring, disaster response, urban planning, climate studies, and even in some military applications European Space Agency (ESA) 2015 . RGB images are used for various applications, from urban mapping of peaceful cities Corbane et al. 2008 to detecting military targets Bandarupally et al. 2020 and conducting reconnaissance operations Wang et al. 2014 . NDVI is used for civilian purposes, such as crop yield estimation Quarmby et al. 1993 and vegetation health monitoring Bento et al. 2020 , as well as for military applications, such as detecting war damage Shelestov et al. 2023 and using field spectroscopy for defence and security (e.g., locating underground structures) Melillos et al. 2016 . NDWI is also used in war damage analysis Vlasova et al. 2023 ; Skyba , and for critical tasks, such as detection of water bodies Özelkan 2020 and flood monitoring Albertini et al. 2022 . These signals hold great importance for both civilian and military applications. However, the overcast cloud barrier limits our ability to use these valuable signals. In contrast, Sentinel-1 European Space Agency (ESA) 2014 operates using Synthetic Aperture Radar (SAR), which is immune to cloud cover and lighting conditions, making it a promising alternative data source. Some attempts to overcome this ‘cloud-barrier’ have been reported in the literature. Kim et al. Kim and Chung 2024 used a modified version of the Brownian Bridge Diffusion model Li et al. 2023 to generate Optical (RGB) images from Very-High-Resolution (VHR) SAR using the MSAW dataset Shermeyer et al. 2020 , which contains pairs of optical and SAR images. Their model introduced the concept of adding the initial SAR image at every step of the diffusion process, effectively conditioning the process on the initial image. Ahmed et al. Sakib Ahmed et al. 2025 on the other hand used a U-Net architecture Ronneberger et al. 2015 to generate NDWI from Sentinel-1 signals. However, neither approach fully utilized the full potential of generating multispectral imagery. We addressed this gap by generating multispectral images from Sentinel-1 to derive optical images, NDWI, and NDVI using a novel flow matching method. This, in the sequel, removes the cloud cover issue with respect to Sentinel-2 images through our model uniquely leveraging the resilience of Sentinel-1 in this context. To accomplish this task, we build upon the broader landscape of generative models. Now, these models typically follow one of two main paradigms. One such method is Generative Adversarial Networks (GANs) Goodfellow et al. 2014 . In this setup, two models compete: one generates outputs similar to the target data, and the other tries to distinguish between real and generated data. However, GANs suffer from training instability among other issues Wiatrak et al. 2019 . Therefore, image-to-image GAN models, such as Pix2Pix Henry et al. 2021 and CycleGAN Zhu et al. 2017 are not the top choice for image-to-image translation tasks. Another category of methods aims to iteratively translate from one distribution to another, such as diffusion models Sohl-Dickstein et al. 2015 and flow matching Lipman et al. 2022 . Traditionally, these methods typically start from a noise distribution and gradually transform it into the target distribution. Notably, during training of these methods, diffusion goes from the target distribution to the noise distribution, and only in the reverse process of inference they do the opposite to get to the target distribution. On the other hand, flow matching moves directly from noise to target distribution in both training and inference. However, for deterministic translation tasks, such as image-to-image translation, starting from the input distribution is more appropriate. This choice avoids unnecessary meandering in the translation process and reduces the number of steps required. Given that using the input distribution and target distribution as the two endpoints is a more efficient strategy, we now consider some of the common iterative models in image-to-image translation. The Brownian Bridge Diffusion Model (BBDM) Li et al. 2023 learns a discrete sequence of steps to transition from the target distribution to the input distribution. During inference, the model reverses this process, subtracting the learned steps back from the input distribution to reach the target distribution. In contrast, Latent Flow Matching Dao et al. 2023 learns a continuous transformation path from the input latent to the target latent distribution. Although these approaches differ in their training direction and granularity (discrete vs. continuous), fundamentally, both aim to model the transformation path between two distributions. One could argue that the training direction is a significant difference, but in essence, both methods are designed to learn this path. We further argue that Flow Matching is a superior approach due to its continuous nature, which allows it to be viewed as a superset of BBDM by covering more intermediate steps during training. This brings us to the concept of the path itself—what exactly defines it? The answer lies in interpolation. In this context, interpolation refers to the scheduling mechanism that guides the model in determining how much to update its state at each step. This “update” quantifies the extent to which the model should translate towards the target distribution at a given time step. Importantly, the early steps in the process tend to be more error-prone. Therefore, assigning equal weight to all steps—as in linear scheduling—is suboptimal, particularly for tasks resembling optimal transport. To mitigate this, smaller updates should be applied in the early stages, motivating the adoption of non-linear scheduling strategies. Among various non-linear scheduling methods, we focus on two specific options: exponential and cosine scheduling. While the exponential and linear schedules were discussed in the original Flow Matching paper Lipman et al. 2022 , to the best of our knowledge, the cosine schedule has not been explored in prior literature of flow matching, including Lipman et al. 2022 ; Lipman et al. 2024 ; Dao et al. 2023 . We propose the cosine scheduling as a novel addition to the methods for flow matching in this context. Cosine scheduling, similar to exponential scheduling, assigns lower weights to the initial steps. However, the update magnitude gradually increases toward the final steps. In other words, the effect of the predicted direction vector from initial distribution towards target distribution becomes more pronounced near the end, allowing for more significant changes as the model approaches the target distribution. As we will later see, each non-linear scheduling has its own benefits. Additionally, we propose a novel modified multi-staged training procedure that specifically addresses the common issue of large errors during the initial stages of translation. Our method places greater emphasis on the early steps and aligns training more closely with the inference process. Furthermore, to stabilize and guide the translation, we condition the model on the initial distribution at every step. 2 Results and Discussion 2.1 Methodical Overview In this study, we transform Sentinel-1 inputs into Sentinel-2 outputs via a learned mapping in a compact latent space. As shown in the Fig. 2(a) process is composed of several key stages. Image Scaling Procedure To standardize input distributions and improve training stability, all images were normalized per channel using a custom scaler named ImageScaler . Scaling was based on percentile thresholds MeVisLab Documentation 2025 ; Colom and Buades 2013 to reduce the influence of outliers. For Sentinel-1, the 0.1 th and 99.9 th percentiles were used; for Sentinel-2, we used the 1 st and 98 th percentiles. Each value was linearly scaled as follows (Eqn. 1 ). No clipping was applied, preserving all values during training. \text{scaled}=\frac{\text{value}-\text{pmin}}{\text{pmax}-\text{pmin}+\varepsilon} (1) Latent Encoding. Sentinel-1 and Sentinel-2 images were encoded using separate Vector-Quantized Variational Autoencoders (VQ-VAE) Van Den Oord et al. 2017 , each reducing spatial resolution by a factor of 2 and projecting the inputs into a 16-channel latent space. This ensures that both 2-channel Sentinel-1 and 4-channel Sentinel-2 inputs are mapped to the same latent dimensionality, enabling computation of the difference vector \Delta Z_{s}=Z_{S_{2}}-Z_{S_{1}} , where Z_{S_{1}} and Z_{S_{2}} denote the latent representations of Sentinel-1 and Sentinel-2, respectively. This difference vector guides the model in learning the transformation from initial to target distribution. Translation Model. The core model is a U-Net architecture Ronneberger et al. 2015 ( UNet2DModel ) that operates entirely within the latent space. Once the images were encoded into their latent forms, we trained the model following Algorithm 1 . The training process included several notable techniques. First, we applied cosine-interpolated mixing: at each training step, we generated an intermediate latent code x_{t} as a weighted blend of Z_{S_{1}} and Z_{S_{2}} , with the weight m_{t}\in[0,1] determined by a cosine schedule (Eq. 4 ). This approach allowed the model to learn smooth transitions from input to output with smaller updates initially and larger steps in the later phases, improving generalization across intermediate representations. To further enrich the training, we employed a multi-stage procedure that incorporated three types of sampling per batch, described as follows. In the first stage of our training, referred to as the continuous mode, m_{t} was uniformly sampled from the interval [0,1] to generate a random x_{t} from the continuous path for each example. In the second stage, referred to as the discrete mode, an integer t\in\{0,\dots,N-1\} was chosen and we set m_{t}=t/N to ensure coverage of fixed points along the interpolation path. Lastly, in the boundary focus mode, we always included m_{t}=0 (i.e., pure Z_{S_{1}} ) to strengthen the model’s learning of the most difficult transformation. Post-processing & Indices. After inference, we used the steps described in Algorithm 2 to decode the predicted latent representation \hat{Z}_{S_{2}} into a reconstructed Sentinel-2 image. We then separated the RGB and NIR channels to compute two indices, namely, the Normalized Difference Water Index (NDWI, Eq. 5 ) and the Normalized Difference Vegetation Index (NDVI, Eq. 6 ). Although we have used a global coverage dataset (see the “ Datasets ” Section in the Supplementary Material , it does not cover every single environment in the world. So, for practical applications on a particular location, we fine-tune the model with some limited cloud-free data of those locations, as shown in Fig. 2(b) . Algorithm 1 Training with Cosine Interpolation from Latent Space of Sentinel-1 to Latent Space of Sentinel-2 1: Input: Number of epochs E , number of interpolation steps N , random noise factor r , data loaders train_latent_loader and val_latent_loader (containing latent embeddings of Sentinel-1 and Sentinel-2), model \mathcal{M} , optimizer, scheduler 2: Initialize: Time step size \Delta t\leftarrow\frac{1}{N} 3: for each epoch e=1 to E do 4: Set model to training mode: \mathcal{M}\leftarrow\texttt{train mode} 5: Initialize total training loss: L_{\text{total}}\leftarrow 0 6: for each batch (\mathbf{z}_{\text{S1}},\mathbf{z}_{\text{S2}}) in train_latent_loader do \triangleright — Training on continuous steps — 7: Sample random timestep: m_{t}\sim\mathcal{U}(0,1) 8: Compute interpolated latent: \mathbf{x}_{t}\leftarrow\frac{1}{2}(1-\cos(\pi m_{t}))\cdot\mathbf{z}_{\text{S2}}+\left(1-\frac{1}{2}(1-\cos(\pi m_{t}))\right)\cdot\mathbf{z}_{\text{S1}} 9: Predict: \hat{\mathbf{y}}\leftarrow\mathcal{M}([\mathbf{x}_{t},\mathbf{z}_{\text{S1}}],m_{t}) 10: Set target: \mathbf{y}\leftarrow\mathbf{z}_{\text{S2}}-\mathbf{z}_{\text{S1}} 11: Compute loss: \mathcal{L}\leftarrow\|\mathbf{y}-\hat{\mathbf{y}}\|^{2} 12: Update model parameters using \mathcal{L} 13: L_{\text{total}}\leftarrow L_{\text{total}}+\mathcal{L} \triangleright — Training on discrete time steps — 14: Sample t\sim\{0,\dots,N-1\},\quad m_{t}\leftarrow\frac{t}{N} 15: Repeat lines 8–13 with the new m_{t} \triangleright — Training at initial step ( t=0 ) — 16: Predict: \hat{\mathbf{y}}\leftarrow\mathcal{M}([\mathbf{z}_{\text{S1}},\mathbf{z}_{\text{S1}}],m_{t}=0) 17: Compute loss: \mathcal{L}\leftarrow\|\mathbf{z}_{\text{S2}}-\mathbf{z}_{\text{S1}}-\hat{\mathbf{y}}\|^{2} 18: Update model parameters 19: L_{\text{total}}\leftarrow L_{\text{total}}+\mathcal{L} 20: end for 21: Compute average training loss: \bar{L}_{\text{train}}\leftarrow\frac{L_{\text{total}}}{\text{num\_batches}} 22: Set model to evaluation mode: \mathcal{M}\leftarrow\texttt{eval mode} , initialize validation loss: L_{\text{val}}\leftarrow 0 23: for each batch (\mathbf{z}_{\text{S1}},\mathbf{z}_{\text{S2}}) in val_latent_loader do 24: Sample t\sim\{0,\dots,N-1\},\quad m_{t}\leftarrow\frac{t}{N} 25: Interpolate: \mathbf{x}_{t}\leftarrow\frac{1}{2}(1-\cos(\pi m_{t}))\cdot\mathbf{z}_{\text{S2}}+\left(1-\frac{1}{2}(1-\cos(\pi m_{t}))\right)\cdot\mathbf{z}_{\text{S1}} 26: Predict: \hat{\mathbf{y}}\leftarrow\mathcal{M}([\mathbf{x}_{t},\mathbf{z}_{\text{S1}}],m_{t}) 27: Compute validation loss: \mathcal{L}\leftarrow\|\mathbf{z}_{\text{S2}}-\mathbf{z}_{\text{S1}}-\hat{\mathbf{y}}\|^{2} 28: L_{\text{val}}\leftarrow L_{\text{val}}+\mathcal{L} 29: end for 30: Compute average validation loss: \bar{L}_{\text{val}}\leftarrow\frac{L_{\text{val}}}{\text{num\_batches}} 31: Step the scheduler using \bar{L}_{\text{val}} 32: if epoch e is a checkpoint interval then 33: Save model checkpoint 34: end if 35: end for Algorithm 2 Inference: Latent-Space Translation with Cosine Schedule 1: Input: pretrained model \mathcal{M} , Sentinel-2 VQ-VAE decoder \mathrm{Dec} , test loader \mathit{test\_latent\_loader} , steps T 2: Output: reconstructed RGB/NIR images, NDVI, NDWI 3: for each batch (\mathbf{z}_{\mathrm{S1}},\mathbf{z}_{\mathrm{S2}}) in \mathit{test\_latent\_loader} do 4: \mathbf{x}\leftarrow\mathbf{z}_{\mathrm{S1}} \triangleright initialize at S1 latent 5: Compute cosine schedule: s_{i}=\tfrac{1}{2}\bigl(1-\cos(\tfrac{\pi\,i}{T-1})\bigr),\;i=0,\dots,T-1 6: for i=0 to T-2 do 7: m\leftarrow s_{i} 8: \delta\leftarrow s_{i+1}-s_{i} 9: \Delta\mathbf{z}\leftarrow\mathcal{M}\bigl([\mathbf{x},\,\mathbf{z}_{\mathrm{S1}}],\,m\bigr) 10: \mathbf{x}\leftarrow\mathbf{x}+\Delta\mathbf{z}.\mathrm{sample}\times\delta 11: end for 12: Decode: \hat{\mathbf{X}}\leftarrow\mathrm{Dec}(\mathbf{x}) \triangleright shape: (B,4,H,W) 13: Split channels: \hat{\mathbf{X}}_{\mathrm{RGB}}=\hat{\mathbf{X}}[:,0:3,:,:] , \hat{\mathbf{X}}_{\mathrm{NIR}}=\hat{\mathbf{X}}[:,3:4,:,:] 14: Compute indices ( \epsilon=10^{-7} ): \mathrm{NDVI}=(\hat{\mathbf{X}}_{\mathrm{NIR}}-\hat{\mathbf{X}}_{\mathrm{RGB}}[:,0:1,:,:])/(\hat{\mathbf{X}}_{\mathrm{NIR}}+\hat{\mathbf{X}}_{\mathrm{RGB}}[:,0:1,:,:]+\epsilon) , \mathrm{NDWI}=(\hat{\mathbf{X}}_{\mathrm{RGB}}[:,1:2,:,:]-\hat{\mathbf{X}}_{\mathrm{NIR}})/(\hat{\mathbf{X}}_{\mathrm{RGB}}[:,1:2,:,:]+\hat{\mathbf{X}}_{\mathrm{NIR}}+\epsilon) 15: Save or return \hat{\mathbf{X}}_{\mathrm{RGB}},\,\hat{\mathbf{X}}_{\mathrm{NIR}},\,\mathrm{NDVI},\,\mathrm{NDWI} 16: end for x_{t}=(1-m_{t})\cdot s_{1}+m_{t}\cdot s_{2},\quad\text{where }m_{t}\in[0,1] (2) x_{t}=\left(1-w_{t}\right)\cdot s_{1}+w_{t}\cdot s_{2},\quad\text{where }w_{t}=\frac{e^{k(m_{t}-1)}-\min}{\max-\min+\varepsilon},\quad m_{t}\in[0,1] (3) x_{t}=\left(1-\frac{1}{2}\left(1-\cos(\pi\cdot m_{t})\right)\right)\cdot s_{1}+\frac{1}{2}\left(1-\cos(\pi\cdot m_{t})\right)\cdot s_{2},\quad\text{where }m_{t}\in[0,1] (4) \text{NDWI}=\frac{\text{Green}-\text{NIR}}{\text{Green}+\text{NIR}} (5) \text{NDVI}=\frac{\text{NIR}-\text{Red}}{\text{NIR}+\text{Red}} (6) Figure 1 : Different Scheme of interpolation or scheduling (a) (b) Figure 2 : a. This subfigure illustrates the training procedure of our CloudBreaker model. We use VQ-VAE latent space representations of cloud-free Sentinel-1 images as input and the corresponding Sentinel-2 latent representations as the ground truth for the U-Net model. After training with the flow matching method, CloudBreaker learns to directly generate Sentinel-2 latent representations from Sentinel-1 inputs. These are then decoded to reconstruct 4-channel images (RGB and Near-Infrared), which are used to compute optical (RGB), NDWI, and NDVI outputs. b. For practical deployment, we fine-tune the model using cloud-free data from the target region. Once fine-tuned, CloudBreaker can be used during real disaster events to generate cloud-free optical products from Sentinel-1 inputs. 2.2 Latent Space results Recall that using two VQ-VAEs, we successfully converted Sentinel-1 and Sentinel-2 images into a latent space. The reconstruction performance from the latent space was evaluated on the decoded outputs. For decoded Sentinel-2, we achieved a mean squared error (MSE) of 0.0006 and a R^{2} score of 0.985. For decoded Sentinel-1, the model achieved an MSE of 0.00054 and a R^{2} score of 0.97. These results indicate that the latent encoding effectively preserved the essential information from the original images, allowing for high-precision reconstruction. 2.3 Translation Results We evaluated our model in two stages: first, in the latent space, where the model learns the mapping (translation) between Sentinel-1 and Sentinel-2 representations; and second, in the decoded image space, where the 4-channel (RGB + NIR) Sentinel-2 output is reconstructed. From the reconstructed images, we further compute and evaluate domain-specific remote sensing indices, such as NDVI and NDWI (which will be shown in the following section). As shown in Table 1 , we compare GAN-based methods, such as Pix2Pix and CycleGAN, diffusion-based methods, such as BBDM, and various configurations of Flow Matching (FM). The evaluation metrics include Fréchet Inception Distance (FID) Dowson and Landau 1982 ; Heusel et al. 2017 , Structural Similarity Index (SSIM) Nilsson and Akenine-Möller 2020 , Learned Perceptual Image Patch Similarity (LPIPS) Zhang et al. 2018 , as well as MSE and R^{2} (see Section S1.3 in the Supplementary Materials). From the results, we observe that FM methods outperform the other approaches by a substantial margin. GAN-based methods generally perform the worst across most metrics, while BBDM comes closer to FM in performance, but still lags behind, particularly in RGB FID (second worst) and RGB LPIPS (worst). Additionally, we comprehensively evaluated the FM model using different schedulers defined in Eq. 2 , Eq. 3 , and Eq. 4 , under two inference step settings (100 and 1000), as summarized in Table 1 . From the radar plot Saary 2008 comparing the top-performing models across multiple metrics in Fig. 3 , we observe that the cosine scheduler (1000 steps) achieves the best perceptual quality, with the lowest FID of 0.6481 , outperforming all other models by more than 50%. Compared to the next best model, the exponential scheduler (Expo k=2 , FID = 0.6840), the cosine scheduler (1000 steps) shows an improvement of approximately 5.25% in FID. While exponential schedulers, such as Expo k=1.5 demonstrate a more balanced performance across certain evaluation metrics— achieving the highest R^{2} score and a low latent-space MSE— the cosine scheduler excels in key structural quality measures, including SSIM of NDVI, SSIM of NDWI, and LPIPS of RGB. This consistent superiority in structural fidelity makes the cosine scheduler the most suitable choice for applications where perceptual realism and visual quality are of primary importance, especially in scenarios involving human interpretation. Additionally, from Table 1 , we also observe that the performance of exponential scheduling is highly sensitive to the parameter k , which controls the steepness of the scheduling curve. Based on metric-specific priorities, users can select models accordingly. Since inference time is a critical consideration for practical applications, we prioritize models that maintain strong performance with fewer inference steps (e.g., 100 steps), as they offer faster generation while still producing realistic images as indicated by lower FID scores. Therefore, our main discussion focuses on cosine interpolation with 100 steps. Figure 3 : Radar plot comparison of top-performing methods on the test dataset, normalized across seven evaluation metrics: FID, MSE_latent, R 2 _latent, RGB SSIM, RGB LPIPS, NDVI SSIM, and NDWI SSIM. The Cosine (1000) scheduler achieves the best perceptual quality with the lowest FID (0.6300) and RGB LPIPS (0.3433), outperforming Expo k=2.5 (1000) by 68.9% in FID and 4.5% in LPIPS. While Expo k=1.5 (1000) yields the highest R 2 _latent (0.6588) and strong NDWI SSIM (0.5886), the Cosine scheduler offers superior perceptual realism. These results indicate a trade-off between perceptual and structural metrics, with the Cosine scheduler emerging as the best choice when visual quality is prioritized. 2.3.1 Translated Sentinel-2 Latent Space Evaluation To assess how effectively the model learns the transformation in the latent space, we compute the Mean Squared Error (MSE) and the coefficient of determination ( R^{2} ) between the predicted latent representation from Sentinel-1 and the ground truth latent embedding of Sentinel-2. The model achieves an MSE of 0.003814 and a R^{2} score of 0.4969 , indicating a reasonable alignment between the predicted and target latent distributions. However, since the R^{2} score overlooks nonlinear, perceptual, and structural factors that are critical for evaluating generated images, it is not the preferred metric for such tasks. Please refer to Section S1.3 in the Supplementary for more details. 2.3.2 Reconstructed Image Evaluation After decoding the latent representation, we obtain the full-resolution 4-channel Sentinel-2 image, consisting of RGB and NIR bands. We evaluated the RGB subset using both perceptual and pixel-wise metrics to comprehensively assess reconstruction quality. The reconstructed RGB images achieved a Fréchet Inception Distance (FID) of 0.7432 , which is significantly low and indicates high perceptual similarity to real Sentinel-2 images—suggesting that the generative model effectively captured the high-level visual features. In terms of pixel-wise evaluation, we observe a mean squared error (MSE) of 0.0198 , which reflects a low average error between reconstructed and ground truth pixel values. The peak signal-to-noise ratio (PSNR) is 17.29 dB , indicating an acceptable level of reconstruction accuracy, especially given the complexity of natural scenes in satellite imagery. The SSIM (range 0 to 1 and the higher the better) is 0.6346 , which shows that essential structural patterns and textures are reasonably well preserved. Additionally, the LPIPS (range 0 to 1 and the lower the better) score of 0.2719 suggests good perceptual similarity at the feature level, as seen by a neural network. Together, these results demonstrate a very high visual fidelity with moderate structural consistency, confirming that the model was able to effectively reconstruct complex natural features. 2.3.3 Remote Sensing Index Evaluation From the reconstructed 4-channel Sentinel-2 image, we computed NDVI and NDWI. These indices are derived from the NIR and visible bands, and help assess vegetation and water body coverage, respectively. Our reconstructed NDVI and NDWI achieved SSIM of 0.6156 and 0.6874 respectively, indicating a good level of structural similarity to the corresponding ground truth indices. SSIM ranges from 0 to 1, where higher values signify closer structural resemblance. For comparison, we evaluated the output of Ahmed et al.’s model Sakib Ahmed et al. 2025 , which only generates NDWI, and obtained an NDWI SSIM of 0.423. This indicates that our model outperforms theirs in preserving the structural features of water-related spectral content in the reconstructed images. The improved SSIM demonstrates better spatial fidelity in our generated NDWI, highlighting the effectiveness of our approach in capturing fine-grained Earth surface characteristics. 2.4 Removing Clouds in Test Set As we can see from Fig. 5 , the use of CloudBreaker has successfully removed the clouds from the test satellite images. The first column shows the optical (RGB) images, while the second and third columns present the corresponding NDVI and NDWI representations for three test samples. Each subfigure presents a pair of images— the left side showing the original cloud-covered image and the right side showing the corresponding image after cloud removal. In the RGB images (Fig. 5 .a, Fig. 5 .d, and Fig. 5 .g), we can now clearly observe underlying surface features and land cover structures that were previously obscured due to cloud cover. This clarity is crucial for accurate visual interpretation and analysis of spatial features. Furthermore, the presence of clouds in the original images negatively affected the quality of the derived vegetation and water indices. The NDVI (shown in Fig. 5 .b, Fig. 5 .e, and Fig. 5 .h) and NDWI (shown in Fig. 5 .c, Fig. 5 .f, and Fig. 5 .i) were distorted due to cloud interference, resulting in incomplete or misleading information. However, after removing the clouds, these indices were properly recovered, revealing the true vegetation health and water distribution in the region. This demonstrates the importance of cloud removal as a preprocessing step for reliable satellite image analysis in environmental monitoring and remote sensing applications. 2.5 Real-life Case Study We have also tested the model for various real-life disastrous events. As it can be observed that Fig. 6 , together with Fig. 7 and Fig. 8 in the supplementary materials, illustrate the optical (RGB), NDVI, and NDWI representations, respectively, of the various disaster events analyzed below. Each figure is organized into five rows corresponding to five different disaster events. Each row contains three columns that represent the satellite imagery before (first column), during (second column), and after (third column) the event. Within each panel (cell), two images are shown side by side: the left image is the original satellite observation (which may include cloud cover, especially in the “during” images, as is common during disasters), and the right image is the corresponding output generated by our model. This structure provides a consistent visual comparison across the different data modalities, highlighting both the impact of each disaster and the effectiveness of the model in reconstructing clearer views under challenging conditions. Below, we present the results of our model in these real-world scenarios. 2.5.1 Amazon Fire Our first real-life example is the Amazon fire that occurred in July 2023. This devastating fire scorched large areas of rainforest, destroyed countless trees, and threatened vulnerable wildlife. It led to severe biodiversity loss, worsened air quality, and released massive carbon emissions. To analyze this event, we used Sentinel-1 imagery from before, during, and after the disaster, removing clouds for better clarity. The first row of Fig. 6 (main text) and Fig. 7 and Fig. 8 (supplementary materials) illustrates the Amazon fire event in optical (RGB), NDVI, and NDWI representations, respectively. Each figure includes the “before” stage: Fig. 6 .a (main text) and Fig. 7 .a and Fig. 8 .a (supplementary materials), where the left column shows the original satellite observation and the right column shows the corresponding output generated by our model. These images clearly depict the intact forest area prior to the fire, with NDVI and NDWI effectively highlighting vegetation health and surface water content, respectively. The b panel — Fig. 6 .b (main text) and Fig. 7 .b and Fig. 8 .b (supplementary materials) — presents the corresponding images during the fire, highlighting the impact on vegetation and water content. The NDVI and NDWI values visibly decline, reflecting the degradation of forest cover and moisture content. Finally, the c panel — Fig. 6 .c (main text) and Fig. 7 .c and Fig. 8 .c (supplementary materials) — shows the area after the fire. A significant reduction in vegetation and water content can be observed in our generated image in the right of those respective figures. These could not be seen in the original images in the left due to being obscured by cloud cover. These visualizations underscore the utility of our model in monitoring and analyzing the progression and aftermath of environmental disasters. 2.5.2 Hurricane Harvey We now turn to our second disaster, Hurricane Harvey hurricane_harvey_wikipedia 2025 , that occurred in the USA in 2017. Hurricane Harvey struck Texas and Louisiana in August 2017, causing catastrophic flooding that displaced over 30,000 people and resulted in more than $125 billion in damages, making it the costliest natural disaster in Texas history . With peak rainfall exceeding 60 inches, it also became the wettest tropical cyclone ever recorded in the United States. The results of applying our model to remove clouds from satellite imagery captured before, during, and after this event are presented below. The second row of Fig. 6 (main text) and Fig. 7 and Fig. 8 (supplementary materials) illustrates the Hurricane Harvey in optical (RGB), NDVI, and NDWI representations, respectively. Fig. 6 .d (main text) and Fig. 7 .d and Fig. 8 .d (supplementary materials) show the optical (RGB), NDVI, and NDWI images of the area before the hurricane. These images clearly delineate the pre-disaster landscape, with NDVI and NDWI effectively capturing healthy vegetation and normal water content. The Fig. 6 .e (main text) and Fig. 7 .e and Fig. 8 .e (supplementary materials) presents the corresponding images during the hurricane, highlighting the disruption to vegetation. NDVI and NDWI values decrease noticeably, reflecting vegetation damage. A noticeable change in land cover can be observed in the generated images on the right, which are not clearly visible in the original images on the left due to cloud cover. Finally, Fig. 6 .f (main text) and Fig. 7 .f and Fig. 8 .f (supplementary materials) show the area after the hurricane. 2.5.3 Nepal Flood Thirdly, we move to the Nepal flooding that happened in 2024 Wikipedia contributors 2024a . Torrential monsoon rains in July, August, and especially late September 2024 caused devastating floods and mudslides across Nepal, severely impacting infrastructure, homes, and agriculture. The floods resulted in over 224 deaths, thousands displaced, widespread damage to bridges, highways, hydropower stations, and disrupted essential services including power and telecommunications. The results from our model for this event are presented in the third row of Fig. 6 (main text) and Fig. 7 and Fig. 8 (supplementary materials). Fig. 6 .g (main text) and Fig. 7 .g and Fig. 8 .g (supplementary materials) present the optical (RGB), NDVI, and NDWI images of the area before the flood. These images clearly delineate the pre-flood landscape, with NDVI and NDWI effectively capturing healthy vegetation and normal water content. Fig. 6 .h (main text) and Fig. 7 .h and Fig. 8 .h (supplementary materials) present the corresponding images during the flood, highlighting the inundation and disruption to vegetation. The NDVI and NDWI values decline noticeably, reflecting the loss of vegetation and increased water coverage. A significant change in land cover and persistent water-logged areas can be observed in the generated images on the right. These details are not visible in the original images on the left due to cloud cover. Finally, Fig. 6 .i (main text) and Fig. 7 .i and Fig. 8 .i (supplementary materials) show the area after the flood. 2.5.4 Cyclone Remal Fourthly, we examine the impact of Cyclone Remal Wikipedia contributors 2024b in 2024. Severe Cyclonic Storm Remal struck West Bengal and Bangladesh in May 2024, causing sustained winds of up to 135 km/h and resulting in at least 85 fatalities. The cyclone disrupted power for around 30 million people in Bangladesh and caused widespread damage in the affected coastal regions. The results from our model for this event are presented in the fourth row of Fig. 6 (main text) and Fig. 7 and Fig. 8 (supplementary materials). Fig. 6 .j (main text) and Fig. 7 .j and Fig. 8 .j (supplementary materials) present the optical (RGB), NDVI, and NDWI images of the area before the cyclone. These images clearly delineate the pre-disaster landscape, with NDVI and NDWI effectively capturing healthy vegetation and normal water content. Fig. 6 .k (main text) and Fig. 7 .k and Fig. 8 .k (supplementary materials) present the corresponding images during the cyclone, highlighting the inundation and disruption to vegetation. NDVI and NDWI values have declined noticeably, reflecting vegetation loss and increased surface water coverage. A significant change in land cover and persistent water-logged areas can be observed in the generated images on the right, which are not clearly visible in the original images on the left due to cloud cover. Finally, Fig. 6 .l (main text) and Fig. 7 .l and Fig. 8 .l (supplementary materials) show the area after the cyclone. 2.5.5 Volcanic Eruption Finally, we examine the impact of the Taal volcanic eruption Wikipedia contributors 2024c in the Philippines in 2020. The Taal Volcano eruption in January 2020 disrupted daily life for millions, causing ashfall across Metro Manila and nearby provinces that worsened air quality and led to widespread evacuations. The event forced residents to leave their homes and affected transportation, health, and local economies in the region. The results obtained from our model for this event are as follows: the first row in .The Fig. 6 .m (main text) and Fig. 7 .m and Fig. 8 .m (supplementary materials) presents the optical (RGB), NDVI, and NDWI images of the area before the eruption. These images clearly depict the pre-eruption landscape, with NDVI and NDWI capturing healthy vegetation and water distribution. Fig. 6 .n (main text) and Fig. 7 .n and Fig. 8 .n (supplementary materials) present the corresponding images during the eruption, highlighting the ash cover, vegetation damage, and alterations in moisture levels. The NDVI and NDWI values decline substantially, reflecting vegetation stress and disruption of surface water patterns. Significant landscape changes can be observed in the generated images on the right, which are not clearly visible in the original images on the left due to heavy cloud and ash cover. Finally, Fig. 6 .o (main text) and Fig. 7 .o and Fig. 8 .o (supplementary materials) show the area after the eruption. Thus, our model was able to generate crucial optical, NDWI, and NDVI representations under different conditions across various disasters. These visualizations demonstrate the utility of our model in monitoring and analyzing the progression and aftermath of various disasters. 2.6 Future Applications Although we have trained CloudBreaker for environments on Earth, its applications need not be limited to our planet. By extending its scope, we can adapt it for use on other celestial bodies, opening the door to uncovering some of the greatest mysteries in space exploration. For instance, it could potentially help us visualize what lies beneath the thick, acidic clouds of Venus by first training on data from various other planets and then fine-tuning on those with similar features. The potential for CloudBreaker in extraterrestrial exploration is immense. It can be applied to other situations where cloud cover is persistent all around like cloud forest. But for this training in similar environment before applying it will be neccessary. We would also suggest to use inference steps of 1000 or higher for better translation of images as this would be uncharted territory. 3 Conclusion We have successfully developed a method capable of removing clouds from Sentinel-2 images by leveraging Sentinel-1 data to reconstruct multi-spectral imagery. These reconstructed multi-spectral images enable the retrieval of optical information as well as the computation of valuable vegetation indices, such as NDWI and NDVI, which are useful for a wide range of practical applications. This work represents a significant step toward overcoming the limitations imposed by cloud cover in remote sensing. Moreover, our novel multi-stage training process facilitates easier learning of the translation path, which can be generalized to other flow matching frameworks. Additionally, our introduction of cosine scheduling into flow matching offers a new tool for use in generative models. Future research can explore alternative non-linear scheduling schemes to further enhance reconstruction quality. Method Steps RGB FID \downarrow MSE {}_{\text{latent}} \downarrow R {}^{2}_{\text{latent}} \uparrow RGB SSIM \uparrow RGB LPIPS \downarrow NDVI SSIM \uparrow NDWI SSIM \uparrow FM-Cosine 100 0.7432 0.003814 0.4969 0.6346 0.2719 0.6156 0.6874 FM-Cosine 1000 0.6481 0.004292 0.4321 0.6193 0.2791 0.5979 0.6722 FM-Expo k=1.0 100 3.0730 0.002453 0.6721 0.5711 0.3361 0.5459 0.5925 FM-Expo k=1.0 1000 2.2845 0.002566 0.6560 0.5619 0.3347 0.5293 0.5812 FM-Expo k=1.5 100 3.2090 0.002488 0.6671 0.5707 0.3321 0.5529 0.5935 FM-Expo k=1.5 1000 2.7888 0.002545 0.6588 0.5670 0.3283 0.5467 0.5886 FM-Expo k=2.0 100 1.4388 0.002620 0.6473 0.5403 0.3447 0.4823 0.5384 FM-Expo k=2.0 1000 0.6840 0.002944 0.6012 0.5104 0.3572 0.4267 0.4912 FM-Expo k=2.5 100 2.3030 0.002613 0.6486 0.5578 0.3321 0.5412 0.5803 FM-Expo k=2.5 1000 2.0311 0.002667 0.6408 0.5541 0.3309 0.5376 0.5760 FM-Linear 100 2.2573 0.002892 0.6077 0.5148 0.3756 0.5062 0.5257 FM-Linear 1000 1.6853 0.003182 0.5661 0.4985 0.3735 0.4900 0.5072 BBDM – 5.9641 0.003352 0.5435 0.4388 0.4286 0.4504 0.5013 Pix2Pix – 3.9371 0.005543 0.2094 0.3160 0.4205 0.2995 0.2948 CycleGAN – 6.9654 0.003898 0.4609 0.4733 0.4278 0.3774 0.4386 Ahmed et al. Sakib Ahmed et al. 2025 – – – – – – – 0.4226 Table 1 : Performance metrics for different schedulers and number of steps. ”FM-” denotes Flow Matching with different interpolation strategies and step counts. ”BBDM” refers to the Brownian Bridge Diffusion Model. ”Pix2Pix” and ”CycleGAN” are supervised and cycle-consistent baselines, respectively, without explicit step-based sampling. Additionally, we evaluated Ahmed et al.’s model Sakib Ahmed et al. 2025 on the NDWI, which is its only output. Figure 4 : a, d, g, j: Comparison between real (left) and generated (right) RGB images. b, e, h, k: Comparison between real (left) and generated (right) NDVI images. c, f, i, l: Comparison between real (left) and generated (right) NDWI images. Figure 5 : a, d, g: Comparison between cloudy real (left) and generated (right) cloud free RGB images. b, e, h: Comparison between cloudy real (left) and generated (right) cloud free NDVI images. c, f, i: Comparison between cloudy real (left) and generated (right) cloud free NDWI images. Figure 6 : Comparison of optical (RGB) images for five real-life disaster events. Each row corresponds to a different disaster: Amazon fire, Hurricane Harvey, Nepal flood, Cyclone Remal, and the Taal Volcano eruption. Each row shows three temporal stages: before (first column), during (second column), and after (third column) the event. Within each cell, the left image is the original satellite observation (often cloud-affected during disasters), and the right image is the corresponding reconstruction generated by our model. a, d, g, j, m : before images; b, e, h, k, n : during images; c, f, i, l, o : after images. 4 Author Contribution Saleh Sakib Ahmed: Significantly contributed to idea formulation, method development, model architecture design, and paper writing. Sara Nowreen: Provided guidance and support, paper writing, and review. M. Sohel Rahman: Provided guidance and support, paper writing, and review. 5