Deep Learning Super Resolution for Satellite Cloud Mask Downscaling

Technical Summary: Deep Learning Super Resolution for Satellite Cloud Mask Downscaling

Problem Statement
Optical satellite data is frequently compromised by haze and clouds, which affect over 67% of the globe, leading to a deficiency in surface information. A persistent challenge in remote sensing is the fundamental trade-off between spatial and temporal resolution; no single sensor can capture data at the highest possible resolution across both dimensions simultaneously. While geostationary satellites like MSG (providing SEVIRI data) offer high temporal resolution, their spatial resolution is limited compared to polar-orbiting satellites like Terra/Aqua (providing MODIS data). This limitation hinders applications requiring continuous, high-resolution cloud monitoring, such as solar energy forecasting, weather prediction, and disaster management. Furthermore, existing cloud mask (CM) products often suffer from misclassifications and a lack of high-resolution imagery for cloud removal applications.

Methodology
The authors propose a supervised deep learning approach to downscale SEVIRI cloud mask products to match the spatial resolution of MODIS, effectively performing a cross-sensor super-resolution (SR) task.

1. Dataset Creation (SEVMOD-CM):

   • Data Sources: The study utilizes SEVIRI data (11 spectral channels + CM product, 3km resolution) from the MSG satellite and MODIS CM products (MOD35/MYD35 L2, 1km resolution) from Terra and Aqua satellites.
   • Preprocessing: Data from 2020–2023 covering Europe and North Africa were temporally synchronized within a 15-minute window.
   • Alignment: SEVIRI data was reprojected to EPSG:4326 to match MODIS. The SEVIRI CM, originally categorical, was binarized and appended as a 12th channel.
   • Patch Generation: Fixed 128×128 MODIS patches were extracted, and corresponding SEVIRI patches were cropped and resampled to a 32×32 grid, creating a 4× upscaling ratio.
   • Final Dataset: From an initial set of ~63,000 pairs, 20,828 pairs containing cloud activity were retained for training and testing (80/10/10 split).

2. Deep Learning Models:
   Two distinct architectures were developed and trained on the SEVMOD-CM dataset:

   • SpatialCNN: A Convolutional Neural Network extending traditional SR approaches with residual learning and progressive upsampling. It accepts a flexible number of SEVIRI input channels (1–12) and uses a sigmoid activation with Binary Cross-Entropy (BCE) loss to output probabilistic cloud masks.
   • SpatialGAN: A Generative Adversarial Network based on SRGAN, adapted for binary cloud mask reconstruction. It employs a composite loss function combining BCE (for pixel-wise classification consistency) with perceptual and adversarial losses (weighted as regularizers for spatial coherence).

3. Evaluation:
   The models were compared against a standard Bicubic Interpolation baseline. Performance was quantified using Peak Signal-to-Noise Ratio (PSNR), Mean Squared Error (MSE), and Structural Similarity Index (SSIM).

Key Contributions

• Novel Dataset: The introduction of SEVMOD-CM, a comprehensive, AI-ready, cross-sensor cloud mask dataset pairing SEVIRI and MODIS observations.
• Dual-Model Framework: The implementation of two advanced SR algorithms (SpatialCNN and SpatialGAN) specifically tailored for the binary nature of cloud mask products and the discrepancies between geostationary and polar-orbiting sensors.
• Cross-Sensor Application: This work represents an initial effort to apply deep learning SR methods directly to the cloud mask products of different satellite sensors, moving beyond single-sensor SR tasks.

Results
Experimental results on the held-out test set (20,828 pairs) yielded the following:

• SpatialGAN achieved the highest Structural Similarity Index (SSIM: 0.7811) and the lowest Mean Squared Error (MSE: 0.0480), indicating superior structural reconstruction and lower error compared to both the CNN and the baseline.
• SpatialCNN performed similarly but slightly worse than the Bicubic Interpolation baseline across all three evaluation metrics (PSNR: 17.99 dB, SSIM: 0.534, MSE: 0.173).
• Bicubic Interpolation achieved the highest PSNR (18.32 dB) but lower SSIM (0.578) and higher MSE (0.12) than SpatialGAN.
• Visual Analysis: Qualitative assessment (Figure 4) demonstrated that SpatialGAN produced sharper, more perceptually accurate reconstructions of cloud shapes and distributions compared to the baseline, despite the inherent complexity of the target MODIS patterns.

Significance and Claims
The paper claims that the proposed models and dataset offer significant value to the remote sensing community by demonstrating the feasibility of applying super-resolution techniques to geostationary-derived cloud mask products. The authors highlight that SpatialGAN, in particular, successfully learns to reconstruct overall cloud structure and distribution, offering improved spatial analysis for applications such as atmospheric monitoring, weather forecasting, disaster risk reduction, solar energy forecasting, and climate research.

The authors modestly acknowledge limitations, noting that inherent spatial, temporal, and spectral misalignments in cross-sensor imagery remain a fundamental challenge that current upsampling methods cannot fully resolve. They suggest future research directions involving advanced architectures like Vision Transformers and diffusion models to further address these cross-sensor SR challenges.