Statistical Descriptors of β-Variational Autoencoder Latent Feature Maps for Image-Based Blueberry Canopy-Size Phenotyping

Plant-level canopy-size assessment is important for precision agriculture applications, including crop monitoring and variable-rate fertilization, but trait-specific annotation remains costly. This study investigates whether unsupervised β-Variational Autoencoder (β-VAE) latent maps from an image reconstruction model encode canopy-size information recoverable through compact statistical descriptors. A β-VAE was trained on the designated training partitions of a dataset of 2043 field-acquired RGB images of individual blueberry plants, without canopy-size labels. The encoder generated spatial posterior mean (μ) and log-variance (logσ2) maps, from which per-location Kullback–Leibler (KL) divergence maps were also computed. These maps were summarized into five descriptor configurations and evaluated using a separate classifier. The 300-feature reference configuration, which combined μ, logσ2, and KL descriptors, achieved 73.11% accuracy, 73.54% balanced accuracy, and 73.42% macro-F1 under a stratified-random split. Independent retraining of the complete pipeline for three acquisition session-held-out evaluations yielded 63.04 ± 4.54% balanced accuracy and 60.92 ± 6.68% macro-F1. Corresponding balanced accuracies were 78.77 ± 7.24% for ImageNet-pretrained EfficientNet-B0 and 54.13 ± 4.95% for a radial basis function support vector machine using 22 image descriptors; summaries are unweighted means ± sample standard deviations across sessions. These findings support a compact and modular framework for image-based blueberry canopy-size phenotyping, in which unsupervised representation learning is separated from downstream supervised classification, although performance remained below supervised transfer learning. The reference trait measures image-relative canopy extent, not physically calibrated plant size.