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

Objective. Plant Positron Emission Tomography (PET) is a new and efficient imaging technique which aims at providing a quantitative analysis of plant stress, enabling personalized crop management and maximizing productivity. However, a highly performant classification system for noisy dynamic plant PET images faces the challenge of retrieving noise-free datasets and encoding both spatial and temporal representations within a unified model.Approach. To overcome these limitations, we introduce an innovative hybrid model that combines denoising and classification for dynamic plant PET images. Initially, we compute a precise solution for the denoising problem of noisy dynamic plant PET images using a modified optimization method coupled with deep convolutional neural networks. Subsequently, this solution is unfolded into a deep network known as the predictor-corrector network (PCNet). To optimize the PCNet without requiring a noise-free dynamic training set, we propose a novel unsupervised learning method. Finally, the sequence of noise-reduced dynamic plant PET images is further fed into a unique classification system, encoding spatial representations of images and temporal representations of multivariate time series into a unified spatiotemporal representation and generating a prediction.Main results. The experimental results underscore the necessity of the denoising procedure and highlight the superiority of the proposed PCNet over existing competing denoising methods, demonstrating the effectiveness of the proposed classification system. Notably, the classification performance between the two classes achieves an averaged accuracy of 0.852, an averaged precision of 0.838, an averaged recall of 0.959, and an averaged F1-score of 0.880.Significance. The ability of the proposed method to reduce noise intensity and effectively encode spatiotemporal representations overcomes the limitations of existing methods. This advancement may have substantial implications for other noisy dynamic image classification.