Source-linked AI summary
Deep Learning Super Resolution for Satellite Cloud Mask Downscaling
Angelos Georgakis, Valentina Kanaki, Giorgos Giannopoulos, Stella Girtsou, Ioannis Kontogiorgakis, Charalampos Kontoes, Kostas Philippopoulos
TL;DR
High-resolution cloud-mask observations remain difficult because satellite sensing faces a persistent spatial–temporal trade-off and cross-sensor cloud-mask super-resolution is largely unexplored. This paper introduces the SEVMOD-CM dataset and CNN- and GAN-based downscaling models, with SpatialGAN achieving higher structural similarity and lower MSE than bicubic interpolation despite slightly lower PSNR.
Problem
Accurate cloud masks with both high spatial and temporal resolution remain limited by satellite sensing trade-offs, misclassifications, and scarce high-resolution imagery.
Method
The paper creates the cross-sensor SEVMOD-CM dataset and trains CNN- and GAN-based models to map SEVIRI inputs to corresponding MODIS cloud masks.
Results
SpatialGAN produced higher structural similarity and lower MSE than bicubic interpolation, while achieving slightly lower PSNR.
Takeaways & Limitations
The findings support geostationary-observation super-resolution as a potentially valuable tool for near-real-time cloud monitoring.
Takeaways & Limitations
Inherent spatial, temporal, and spectral misalignment between cross-sensor imagery remains a fundamental challenge that upsampling methods cannot fully resolve.
Abstract
from arXiv · showhide
A vast amount of optical satellite data is being transmitted to Earth-based servers every day, and more than half of this data is affected by haze or clouds. Additionally, this data suffers from the fundamental trade-off between spatial and temporal resolution, which remains largely unresolved, making the acquisition of continuous high-resolution satellite observations of clouds an ongoing challenge. This work addresses this challenge by proposing two Deep Learning super-resolution methods for the accurate downscaling of SEVIRI cloud mask products, as well as a novel cross-sensor cloud mask dataset called SEVMOD-CM, created by spatially and temporally matching MODIS and SEVIRI satellite observations. The two proposed models are a CNN-based (SpatialCNN) and a GAN-based (SpatialGAN) Neural Network. Trained on the SEVIRI spectral and cloud mask products, the proposed methods predict the corresponding MODIS Cloud masks, achieving a 4x spatial enhancement across sensor domains. Both approaches are evaluated experimentally, and compared against the standard bicubic interpolation upsampling technique. The experimental results demonstrate the value of the proposed models and dataset for the remote sensing community, highlighting the benefits of applying super-resolution techniques to geostationary-derived cloud mask products for applications such as atmospheric monitoring, weather forecasting, disaster risk reduction, solar energy forecasting, and climate research.
I. INTRODUCTION
The introduction motivates accurate, high-resolution cloud-mask products by persistent spatial–temporal resolution trade-offs, cloud misclassifications, and limited high-resolution imagery. It presents SpatialCNN, SpatialGAN, and the SEVMOD-CM dataset as a cross-sensor cloud-mask super-resolution contribution.
- Motivation: Approximately 67% of the globe is covered by haze or clouds, complicating optical remote-sensing retrievals and cloud-mask-dependent applications.The introduction links this atmospheric coverage to the broader need for improved remote-sensing observations.
- Motivation: The spatial–temporal resolution trade-off, cloud misclassifications, and limited high-resolution imagery create a need for accurate high-resolution cloud-mask products.No sensor can capture the highest possible spatial and temporal resolution across all wavebands, while existing products contain misclassifications.
- Background: Super-resolution encompasses interpolation-based, reconstruction-based, and learning-based upsampling techniques, including widely adopted bicubic interpolation.The introduction places super-resolution within computer vision and image processing and identifies bicubic interpolation as a common method.
- Contributions: The paper introduces SpatialCNN, SpatialGAN, and the SEVMOD-CM dataset for the largely unexplored cross-sensor cloud-mask super-resolution task.The models represent CNN-based and GAN-based deep-learning super-resolution families, with task-specific architectural extensions and enhancements.
- Contributions: The study improves spatial analysis of the SEVIRI cloud-mask product, creates an AI-ready cross-sensor dataset, and evaluates the algorithms’ limitations and future directions.The proposed algorithms target the product generated by the SEVIRI sensor aboard the MSG geostationary satellite.
II. MATERIALS AND METHODS
The materials and methods establish terminology for resolution changes and organize the workflow into data acquisition, SEVMOD-CM creation, model selection, and super-resolution evaluation.
- Terminology: The paper uses “downscale” for transitioning from low to high resolution in Earth observation and remote sensing, whereas computer vision commonly uses “upscale” for the same process.“Upsampling” and “downsampling” are also used as alternatives to upscaling and downscaling.
- Workflow: The methodology begins by finding and downloading the necessary data.
- Workflow: The workflow then performs data postprocessing and creates the SEVMOD-CM dataset.
- Workflow: Subsequent stages select and implement super-resolution deep-learning models, then evaluate the super-resolution results.The paper presents these stages in more detail and provides a roadmap for dataset creation.
A. SEVMOD-CM DATASET · 1) DATA ACQUISITION:
The SEVMOD-CM dataset was developed from paired low- and high-resolution observations collected by SEVIRI on MSG and MODIS on Aqua and Terra. The data include 11 SEVIRI spectral channels, a 3 km SEVIRI cloud-mask product, and a 1 km MODIS cloud-mask product spanning 2020–2023.
- 1) DATA ACQUISITION:: SEVIRI observations were collected from the instrument onboard Meteosat Second Generation, while MODIS observations came from the Aqua and Terra satellites.The dataset pairs low- and high-resolution satellite observations across these sensors.
- 1) DATA ACQUISITION:: 11 SEVIRI spectral channels and the 3 km cloud-mask product were collected alongside the 1 km MODIS cloud-mask product.The SEVIRI cloud-mask product was separately downloaded and later processed in a combined netcdf file.
- 1) DATA ACQUISITION:: 2020–2023 observations were used to develop the SEVMOD-CM dataset.The supplied passage states that the observations covered Euro…
2) DATA PROCESSING:
The data-processing pipeline converted downloaded MODIS and SEVIRI products into paired image patches that were temporally synchronized and spatially aligned for deep-learning super-resolution training. MODIS cloud-mask data were decoded from 48-bit ancillary flags and test results into a binary cloud mask, while SEVIRI cloud masks initially contained categorical values.
- Paired image preparation: Downloaded MODIS and SEVIRI products were preprocessed into paired image patches that were temporally synchronized and spatially aligned for deep-learning super-resolution training.The resulting patches were intended to support training of DL SR algorithms.
- MODIS cloud-mask decoding: 48-bit MODIS cloud-mask ancillary flags and test results were decoded using bitwise operators and converted into a binary cloud mask.The MODIS CM product initially encoded a large set of ancillary flags and test results in 48-bit representation.
- SEVIRI cloud-mask processing: The SEVIRI cloud mask initially contained categorical values.The supplied passage begins listing the categorical encoding but is truncated.
B. Super Resolution Deep Learning Models
The paper formulates cloud-mask super-resolution as supervised cross-sensor mapping from low-resolution SEVIRI inputs to corresponding high-resolution MODIS masks. It develops SpatialCNN and SpatialGAN architectures tailored to this binary cloud-mask task.
- Model formulation: The supervised task maps LR multi-spectral SEVIRI radiances and cloud masks to spatially and temporally corresponding HR MODIS cloud masks.This formulation addresses cross-sensor and cross-resolution discrepancies between geostationary SEVIRI and polar-orbiting MODIS observations.
- SpatialCNN: SpatialCNN extends traditional CNN super-resolution approaches with residual learning and progressive upsampling.The model is described as a deep CNN architecture for the cross-sensor cloud-mask super-resolution task.
- SpatialGAN: SpatialGAN adapts SRGAN to binary cloud-mask super-resolution and supports a flexible number of SEVIRI input channels.Its composite loss was modified because SRGAN was originally designed for photo-realistic image reconstruction, rather than binary cloud masks.
III. RESULTS AND DISCUSSION
The work produced the SEVMOD-CM cross-sensor dataset and evaluated SpatialCNN and SpatialGAN for cloud-mask super-resolution. SpatialGAN generated sharper, more perceptually accurate reconstructions than SpatialCNN and improved structural similarity over bicubic interpolation, despite unresolved cross-sensor misalignment challenges.
- Dataset: The SEVMOD-CM Dataset contains 4.8GB across 126216 .tif files, representing 63108 complete MODIS-SEVIRI pairs before cloud-free-pair filtering.SEVIRI patches contain 12 bands, while MODIS provides the binary cloud-mask target.
- Model comparison: SpatialGAN produced sharper and more perceptually accurate cloud-mask reconstructions than SpatialCNN.The results attribute this improvement to SpatialGAN’s ability to reconstruct overall cloud structure and distribution.
- Model comparison: Compared with bicubic interpolation, SpatialGAN achieved slightly lower PSNR but significantly higher SSIM and lower MSE.The result suggests reduced high-frequency-detail recovery alongside improved structural similarity.
- Limitations: Cross-sensor spatial, temporal, and spectral misalignment remains a fundamental challenge that image upsampling methods cannot yet fully resolve.Complex and scattered MODIS cloud patterns also contribute to lower numerical metrics.
IV. CONCLUSION
The paper presents a cross-sensor satellite cloud-mask super-resolution approach using SEVIRI and MODIS imagery, two deep-learning models, and a task-specific dataset. Selected configurations outperform the baseline in structural similarity and reconstruction accuracy, underscoring GAN effectiveness for this task.
- Contributions: The study develops two deep-learning super-resolution models and a task-specific dataset for cross-sensor satellite cloud-mask super-resolution using SEVIRI and MODIS imagery.The paper also demonstrates their applicability to the task.
- Results: Selected model configurations achieve much higher structural similarity and reconstruction accuracy than the baseline method.The passage does not specify numerical metric values.
- Implications: The findings highlight the effectiveness of GAN-based models for cross-sensor satellite cloud-mask super-resolution.This conclusion follows from the reported gains of specific model configurations over the baseline.