On the value of CTIS imagery for neural-network-based classification: a simulation perspective

Area of Science:

• Hyperspectral Imaging
• Machine Learning in Agriculture
• Computational Imaging

Background:

• Computed tomography imaging spectrometers (CTIS) capture multiplexed spatiospectral projections, traditionally requiring 3D data reconstruction.
• Hyperspectral imaging is valuable in agriculture for tasks like disease detection.
• Current CTIS analysis relies on reconstructing the full hyperspectral cube.

Purpose of the Study:

• To investigate the feasibility of learning information directly from raw CTIS output.
• To apply compressed learning for binary classification using CTIS data.
• To develop and utilize a novel CTIS simulator for agricultural applications, specifically apple scab detection.

Main Methods:

• Training a neural network to perform binary classification directly on raw CTIS images.
• Developing a novel CTIS simulator that preserves realistic pixel intensities.
• Simulating CTIS images of apple leaves with varying degrees of scab infection.

Main Results:

• The trained neural network successfully performed binary classification on simulated CTIS images.
• Demonstrated the potential for direct information extraction from CTIS raw output.
• Validated the utility of the novel CTIS simulator for generating realistic training data.

Conclusions:

• Compressed learning is a viable approach for analyzing CTIS data, eliminating the need for 3D reconstruction.
• This method shows promise for efficient and direct disease detection in agriculture.
• The developed CTIS simulator facilitates research in hyperspectral imaging applications.