A hybrid predictor-corrector network and spatiotemporal classifier method for noisy plant PET image classification

Area of Science:

• Agricultural Imaging
• Biophysics
• Machine Learning

Background:

• Plant Positron Emission Tomography (PET) offers quantitative plant stress analysis for personalized crop management.
• Noisy dynamic plant PET images present challenges for classification due to noise and the need for unified spatiotemporal representation.

Purpose of the Study:

• To develop an innovative hybrid model for denoising and classifying dynamic plant PET images.
• To address limitations in retrieving noise-free datasets and encoding spatiotemporal information.

Main Methods:

• A modified optimization method with deep convolutional neural networks was used for denoising.
• A predictor-corrector network (PCNet) was developed and optimized using unsupervised learning.
• A classification system was designed to unify spatial and temporal representations into a spatiotemporal format.

Main Results:

• The proposed PCNet demonstrated superiority over existing denoising methods.
• The classification system achieved an average accuracy of 0.852, precision of 0.838, recall of 0.959, and F1-score of 0.880.
• The denoising procedure was shown to be essential for effective classification.

Conclusions:

• The hybrid model effectively reduces noise and encodes spatiotemporal information in dynamic plant PET images.
• This advancement has significant implications for noisy dynamic image classification beyond plant science.