Lightweight cloud masking models for on-board inference in hyperspectral imaging.

Area of Science:

• Remote Sensing
• Artificial Intelligence
• Machine Learning

Background:

• Cloud and cloud shadow masking is essential for processing hyperspectral satellite imagery.
• Accurate data preprocessing is vital for reliable space-based applications.

Purpose of the Study:

• To investigate lightweight machine learning models for cloud and cloud shadow masking in hyperspectral satellite imaging.
• To identify models suitable for deployment on satellites with limited resources.

Main Methods:

• Evaluated gradient boosting models (XGBoost, LightGBM) and convolutional neural networks (CNNs).
• Focused on models with reduced resource requirements and efficient deployment capabilities.
• Assessed model performance based on accuracy, storage, and inference speed.

Main Results:

• All evaluated boosting and CNN models achieved over 93% accuracy.
• A CNN with feature reduction demonstrated superior efficiency, balancing high accuracy with low storage and fast inference.
• Optimized CNN variations with minimal trainable parameters (under 597) showed the best trade-off for deployment.

Conclusions:

• Lightweight AI models, particularly optimized CNNs, show significant potential for real-time hyperspectral image processing.
• These models can support the development of efficient on-board satellite AI systems.
• The findings facilitate the creation of analysis-ready hyperspectral data for diverse space-based applications.