Impact of various factors on partial least squares model robustness for nondestructive peach fruit quality assessment
Date
2023
Journal Title
Journal ISSN
Volume Title
Abstract
Given declining fruit consumption due to poor fruit quality and large amounts of waste, peach growers have continuously suffered from financial loss and the industry has seen a sharp decline in recent decades. Due to the time consuming and destructive nature of conventional fruit quality assessment, many peach growers prioritize fruit characteristics conducive to shipping and storage over characteristics which correlate with consumer acceptance. This prioritization has resulted in the poor-quality fruit which consumers have grown to associate with fresh peaches and contributed to large annual waste. A potential solution is the use of near-infrared spectroscopy (Vis-NIRS) paired with partial least squares (PLS) modeling, as a field deployable method that can be used to measure preharvest internal fruit quality to produce information quickly and non-destructively. These qualities offer an answer to declining fruit quality and waste. Although promising, the technology is only as good as the data used to train the models. Quality data is hard to collect as it requires the consideration of many factors including the temperature of the sample and the inclusion of biological variability impacted by seasonal changes, cultivar differences, fruit maturity, and many management factors such as crop load, rootstocks, irrigation regimes, and training systems to capture the relationships needed for good model performance. In tree fruit research, handheld Vis-NIRS devices have been used to predict internal quality parameters such as sweetness (dry matter content, DMC; soluble solids concentration, SSC) and fruit physiological maturity related to chlorophyll content (index of absorbance difference, IAD). Although accurate, the statistical models used to make such predictions often struggle with robustness across cultivars and growing seasons and regions due to a lack of biological variability, or a lack of representative data from factors like temperature. These challenges have led to slow industry adoption. To address this issue, models for 13 distinct peach cultivars were constructed by combining data from two seasons (2016 and 2021) followed by external validation with data from a third season (2022). The data from 2016 was collected over a range of preharvest factors, fruit development stages and temperatures, and the inclusion of 2021 data added additional biological variability. External validation produced error rates of 0.36 - 0.42%, 0.59 - 0.63%, and 0.05 - 0.04 for DMC, SSC and IAD, respectively, across the 13 peach cultivars indicating the models trained in 2021 were robust and performing at an acceptable level to impact grower decision making. It was observed that the additional inclusion of data from different cultivars and growing environments, as well as a third growing season (2017) did not significantly impact model performance. The lack of improvement suggests that the data from each year contain enough covariate variability to cover a broad range of measurements (i.e. input values) that growers and researchers are likely to observe when collecting data to predict peach quality in different orchards or seasons. This insensitivity to various environmental and growing conditions, generally referred to as external factors, due to the variability captured in the data used to build model is characteristic of a robust model.
Description
Rights Access
Subject
PLS
Vis-NIRS
quality
peach