Stem water potential estimation from images using a field noise-robust deep regression-based approach in peach trees

Abstract

Field-grown peach trees are large and have a complex branch structure; therefore, detection of water deficit stress from images is challenging. We obtained large datasets of images of field-grown peach trees with continuous values of stem water potential (Ψstem) through partial secession treatment of the base of branches to change the water status of the branches. The total number of images as frames extracted from videos of branches was 23,181, 6743, and 10,752, in the training, validation, and test datasets, respectively. These datasets enabled us to precisely model water deficit stress using a deep-learning-regression model. The predicted Ψstem of frames belonging to a single branch showed a Gaussian distribution, and the coefficient of determination between the measured and predicted values of Ψstem increased to 0.927 by averaging the predicted values of the frames in each video. This method of averaging the predicted values of frames in each video can automatically eliminate noise and summarize data into the representative value of a tree and is considered to be robust for the diagnosis of water deficit stress in large field-grown peach trees with a complex branch structure.