MODAMS: design of a multimodal object-detection based augmentation model for satellite image sets.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral satellite image classification requires efficient augmentation techniques.
• Existing models lack adaptability and image-specific parameter consideration, limiting efficiency.
• Dynamic band fusions and static deep learning augmentations are insufficient.

Purpose of the Study:

• To propose a novel multimodal object-detection based augmentation model for hyperspectral satellite images.
• To enhance classification performance by optimizing band selection and augmentation strategies.
• To overcome limitations of existing static and non-adaptive augmentation methods.

Main Methods:

• Customized YOLO (You Only Look Once) object detection on hyperspectral bands.
• Cascaded dual Generative Adversarial Network (cdGAN) for object-level importance estimation.
• Elephant Herding Optimization (EHO) for hyperspectral band selection and Firefly Optimizer (FFO) for augmentation control.
• Incremental Learning (IL) layer for continuous accuracy improvement.

Main Results:

• Improved classification accuracy by 8.5%.
• Enhanced precision by 4.3% and recall by 6.5%.
• Reduced classification delay by 2.9% compared to existing methods.

Conclusions:

• The proposed multimodal object-detection model significantly enhances hyperspectral satellite image classification.
• The integration of EHO, FFO, and IL optimizes augmentation and band selection for improved performance.
• This approach offers a more efficient and accurate solution for satellite image analysis across various applications.