What the study found
Lightweight machine learning models for cloud and cloud shadow masking in hyperspectral satellite imaging reached accuracies above 93%. The CNN with feature reduction was the most efficient model in the study, combining high accuracy, low storage needs, and fast inference.
Why the authors say this matters
The authors say these results show the potential of lightweight artificial intelligence models for real-time hyperspectral image processing. They conclude that this supports the development of on-board satellite AI systems for space-based applications.
What the researchers tested
The researchers evaluated several lightweight machine learning approaches for hyperspectral cloud and cloud shadow masking. The models included gradient boosting methods such as XGBoost and LightGBM, as well as convolutional neural networks (CNNs).
What worked and what didn't
All of the boosting and CNN models achieved accuracies exceeding 93%. Among the models studied, the CNN with feature reduction offered the best balance of deployment feasibility, accuracy, computational efficiency, and rapid inference on both CPUs and GPUs. Variations of this CNN had only up to 597 trainable parameters and were reported as the best trade-off.
What to keep in mind
The abstract does not describe limitations or caveats beyond the focus on lightweight models for on-board inference. The summary available here does not provide details on datasets, evaluation settings, or performance differences beyond the reported accuracy and efficiency comparisons.
Key points
- Cloud and cloud shadow masking is described as an important preprocessing step for hyperspectral satellite imaging.
- All tested boosting and CNN models achieved accuracies above 93%.
- The CNN with feature reduction was reported as the most efficient model.
- Some versions of the feature-reduced CNN had up to 597 trainable parameters.
- The authors say the results support development of on-board satellite AI systems.
Disclosure
- Research title:
- Lightweight models achieved strong cloud masking accuracy
- Authors:
- Mazen Ali, António B. Pereira, Fabio Gentile, Aser Cortines, Sam Mugel, Román Orús, Stelios P. Neophytides, Michalis Mavrovouniotis
- Institutions:
- Multiverse Computing (Spain), Cyprus University of Technology
- Publication date:
- 2026-04-22
- OpenAlex record:
- View
Get the weekly research newsletter
Stay current with peer-reviewed research without reading academic papers — one filtered digest, every Friday.