UC Davis

Nondestructive imaging combined with recent advances in data processing techniques allows for minimally invasive, high-throughput, time-series analysis in both food and agriculture. The flexibility of neural network architectures allows for prediction of continuous output metrics, which may be valuable in analysis of both food and agriculture. In this work, nondestructive image analysis was used to demonstrate the relevance of image data to physical properties in food and agricultural systems as well as the potential for predictive models to increase the efficiency of quantitative data analysis.Apples undergoing in vitro gastric digestion were used as a model food system, and X-ray micro-computed tomography data were shown to relate to peak force of apple tissue during compression. This relationship was probed further using a deep learning approach, where the compression curves of apple tissue during in vitro gastric digestion were predicted using a regression convolutional neural network (CNN) model. Results under cross-validation demonstrated strong accordance between predicted and measured compression curves, with an R2 of 0.939 and RMSE of 4.36 N across measured force values. This relationship declined in samples from a holdout set, with an RMSE of 14.3 N, although this result was influenced strongly by the incubation medium tested (water vs. gastric juice). Within the agricultural domain, the task of nondestructive yield estimation was demonstrated in vineyards. Images of grapevines collected using proximal sensing were associated with yield measured at harvest using a commercial yield monitor, allowing for a dataset of 23,581 yield measurements predicted using 164,699 images to be collected. Three deep learning architectures were used to predict the measured yield values from the images: object detection, CNN regression, and a transformer regression network. Regression-based architectures were used to eliminate the bottleneck of hand-labeling images. Results demonstrated that regression methods performed comparably to object detection methods without the need for hand labeling. Grape yield from within a training set, was correlated with model output at a mean absolute percent error of 6.4% when aggregated into 10 m regions using a transformer model. This study also demonstrated performance on a representative holdout set, with an error of 18% obtained using the same model and conditions. Overall, this study demonstrated the potential of deep learning combined with nondestructive imaging for quantitative analysis of food and agricultural systems. Results demonstrated that image data can be related to mechanical properties of food materials as well as yield in an agricultural environment.