📄 agriculture

Leave-One-Day-Out Cross-Validation of NIR-PLSR Models for Simultaneous Prediction of Eight Banana Quality Attributes during 12-Day Storage

This study demonstrates that near-infrared spectroscopy combined with partial least squares regression and leave-one-day-out cross-validation enables the simultaneous, non-destructive prediction of multiple banana quality attributes during 12-day storage, revealing that while models for total soluble solids, monosaccharides, and firmness perform well, temporal generalization is often overestimated by traditional random cross-validation methods.

Original authors: Pu Zhang, Feng Zhang, Jin Luo, Wenlin Wang, Xiaohe Wang, Hanxia Liu, Fang Lin, Hong Zhao, Jinfeng Peng

Published 2026-09-10

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Pu Zhang, Feng Zhang, Jin Luo, Wenlin Wang, Xiaohe Wang, Hanxia Liu, Fang Lin, Hong Zhao, Jinfeng Peng

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Bananas are a study in contradictions. They arrive at the market hard, green, and starchy, yet they possess an internal clock that drives them to soften, sweeten, and eventually rot. This transformation is not merely a change in appearance; it is a complex biochemical event where stored starch breaks down into sugars, cell walls dissolve, and acidity shifts. For farmers and retailers, knowing exactly where a banana sits on this timeline is crucial. If a fruit is picked too early, it never ripens properly; if it is picked too late, it spoils before it reaches the consumer. Traditionally, the only way to know the internal state of a banana is to cut it open, crush the flesh, and run it through chemical tests. This method destroys the fruit, making it impossible to track the same banana as it ripens day by day.

To solve this, scientists have turned to light. Near-infrared spectroscopy is a technique that uses invisible light to peer inside fruit without touching it. When this light hits a banana, it bounces off in a pattern that changes depending on what is inside—how much water is present, how much sugar has formed, and how firm the tissue remains. By analyzing these light patterns, researchers hope to build a model that can predict the quality of the fruit instantly and non-destructively. However, building such a model is tricky because bananas change rapidly. A banana from yesterday looks very different from one tomorrow, and a model trained on a mix of days might learn to recognize the date rather than the fruit itself.

A team of researchers from Xi'an Jiaotong University and the China National Research Center for Food Safety Inspection and Testing set out to test this idea with a rigorous approach. They gathered seventy-two bananas and stored them in a controlled environment at a steady temperature and humidity for twelve days. Every single day, they measured the same fruit using a near-infrared spectrometer, which captures the light reflection, and then immediately destroyed that specific fruit to measure its actual chemical contents. They tracked eight different qualities: three types of sugar (fructose, glucose, and sucrose), starch, total soluble solids (a measure of sweetness and dissolved matter), firmness, pH, and acidity.

The researchers faced a specific challenge in how they tested their predictions. In many studies, scientists mix all their data together and randomly split it into training and testing groups. This method often gives overly optimistic results because the model can learn from samples that are almost identical to the ones it is supposed to predict. To avoid this, the team used a method called leave-one-day-out cross-validation. Imagine a model that is trained on data from eleven days of storage and then asked to predict the quality of the fruit on the twelfth day, a day it has never seen before. They repeated this process, leaving out each of the twelve days in turn, to see if the model could truly generalize to new, unseen time periods.

The results revealed a clear divide between what the light could see and what it could not. The model was remarkably successful at predicting the attributes that change most dramatically during ripening. It could accurately forecast the levels of glucose and total soluble solids, with a performance so strong that it could distinguish between high and low values with great precision. It also performed well for fructose, starch, and the firmness of the fruit. These are the qualities that define the transition from a hard, starchy green banana to a soft, sweet yellow one. The light patterns captured the massive shift in sugar content and the breakdown of the fruit's structure with high reliability.

However, the model struggled with the more subtle changes. It could only offer rough estimates for sucrose, the specific sugar that often peaks in the middle of the ripening process, and it failed to predict the fruit's acidity and pH levels with any useful accuracy. The researchers found that the changes in acidity were simply too small and too noisy for the light sensors to detect against the background of the fruit's other properties. This suggests that while light is excellent for tracking the major structural and sugar changes, it is not yet sensitive enough to measure the fine chemical shifts in tartness.

A key finding of the study was that the way the data is tested matters more than the technology itself. When the researchers compared their strict "leave-one-day-out" method against the standard random-split method, they found that the random method significantly overestimated how well the models would work in the real world. The random method made the models look better than they actually were because it allowed the model to access similar data from other days. The strict time-based testing showed that while the models are promising, they are not perfect, and their ability to predict the future depends heavily on how they are validated.

To make the technology more practical, the team also looked at whether they could simplify the sensors. Instead of using the entire range of light wavelengths, they identified specific bands of light that carried the most important information. They found that by focusing only on these key wavelengths, they could reduce the complexity of the system by more than half without losing much predictive power. This means that in the future, a simple, low-cost device using just a few specific colors of light could potentially be used to sort bananas on a conveyor belt, telling workers exactly which fruits are ready to eat and which need more time, all without cutting a single one open.

The study concludes that while near-infrared light is a powerful tool for monitoring the ripening of bananas, it is not a magic wand that solves every problem. It excels at tracking the big changes in sugar and texture but has limits when it comes to acidity. The research also serves as a warning to the scientific community: when studying things that change over time, like fruit ripening, you cannot rely on random testing methods. You must test your models against the future they are meant to predict. By doing so, the researchers have provided a realistic roadmap for developing non-destructive tools that can help reduce food waste and ensure that the bananas we eat are at their peak.

Technical Summary: Leave-One-Day-Out Cross-Validation of NIR-PLSR Models for Simultaneous Prediction of Eight Banana Quality Attributes during 12-Day Storage

Problem Statement
Postharvest banana ripening is a dynamic, time-series process involving coordinated biochemical changes, including starch degradation, sugar accumulation, tissue softening, and acidity shifts. Conventional methods for monitoring these attributes (fructose, glucose, sucrose, starch, total soluble solids [TSS], firmness, pH, and titratable acidity [TA]) are destructive and labor-intensive, preventing repeated measurements on the same fruit. While Near-Infrared (NIR) spectroscopy combined with Partial Least Squares Regression (PLSR) offers a non-destructive alternative, existing research often relies on random sample splitting or standard k-fold cross-validation. These methods fail to account for the temporal autocorrelation inherent in ripening data, where samples from adjacent days are spectrally similar. Consequently, standard validation strategies may overestimate model performance by allowing "data leakage" between training and validation sets, leading to models that fail when predicting unseen storage durations. Furthermore, there is a lack of systematic studies simultaneously tracking carbohydrate dynamics and other quality attributes while evaluating time-series generalization using rigorous temporal validation.

Methodology
The study utilized 72 banana samples (one per bunch) stored under controlled conditions (21 ± 1 °C, 85 ± 5% RH) for 12 days. Spectra were acquired in diffuse-reflection mode (1000–1799 nm) using a SupNIR-1500 spectrometer. Reference values for eight attributes were obtained via destructive analysis (HPLC for sugars/starch, texture analyzer for firmness, titration/refractometry for pH/TA/TSS).

The core methodological framework involved:

Spectral Preprocessing: Seven strategies were compared, including Raw, Savitzky-Golay smoothing (SG0) and first derivative (SG1), Standard Normal Variate (SNV), Multiplicative Scatter Correction (MSC), and their combinations with SG1.
Modeling: PLSR models were developed to relate spectral data to physicochemical attributes.
Validation Strategies:
- Repeated k-fold Cross-Validation (CV): Used for parameter optimization (selecting latent variables) within the calibration set.
- Leave-One-Day-Out Cross-Validation (LODO-CV): The primary evaluation metric. In this approach, data from one specific storage day were reserved as the validation set while data from the remaining 11 days formed the calibration set. This was repeated 12 times. This method strictly tests the model's ability to generalize to unseen temporal data.
Variable Selection: Informative wavelengths were selected based on Variable Importance in Projection (VIP) scores (>1) and regression coefficients, interpreted against known NIR absorption bands (O–H and C–H overtones).
Evaluation Metrics: Performance was assessed using $R^2$ , RMSE, MAE, Bias, RPD (Ratio of Performance to Deviation), and RPIQ.

Key Results

Temporal Generalization: LODO-CV consistently yielded lower performance metrics compared to repeated k-fold CV across all attributes. This confirms that random splitting overestimates predictive capability for time-series postharvest data.
Prediction Performance (LODO-CV):
- Excellent/Good: Glucose and TSS models achieved the highest performance ( $R^2_P > 0.90$ , RPD > 3.5). Fructose, starch, and firmness models provided good predictions ( $R^2_P = 0.85–0.88$ , RPD = 2.63–2.99).
- Approximate: Sucrose prediction was approximate ( $R^2_P \approx 0.75$ , RPD $\approx$ 2.05).
- Limited: pH and TA models showed poor predictive ability due to narrow variation ranges and weak spectral signals relative to noise.
Preprocessing Impact: The optimal preprocessing method varied by attribute. For instance, SG0 was optimal for glucose and fructose, SG1 for TSS, and raw spectra for firmness.
Variable Selection: Using VIP-selected wavelengths (reducing dimensionality by 44–80% depending on the attribute) resulted in models with performance comparable to full-spectrum models under LODO-CV. The selected wavelengths corresponded primarily to O–H and C–H absorption bands associated with water, sugars, and starch, confirming the chemical relevance of the selected features.
Physicochemical Trends: The study confirmed the expected ripening trajectory: starch depletion (days 5–8), rapid sugar accumulation (days 6–9), and tissue softening. Spectral absorbance increased over time, particularly in the 1400–1500 nm and 1700–1799 nm regions, reflecting changes in water status and carbohydrate composition.

Significance and Claims
The paper claims that the integration of NIR-PLSR with LODO-CV provides a more realistic and rigorous assessment of model generalization for fruit ripening studies than traditional random-split validation. The study demonstrates that while simultaneous non-destructive prediction of multiple banana quality attributes is feasible, the success varies significantly by attribute, with carbohydrates (glucose, fructose, TSS) and firmness being the most predictable.

Key contributions include:

Methodological Rigor: Highlighting the necessity of time-series validation (LODO-CV) to avoid overestimating model utility in dynamic storage scenarios.
Simultaneous Monitoring: Providing a framework for tracking the complex interplay of starch-to-sugar conversion and texture changes within a single experimental setup.
Practical Application: Demonstrating that dimensionality reduction via VIP selection can simplify sensing systems (requiring fewer wavelengths) without sacrificing temporal generalization performance, particularly for TSS, monosaccharides, and firmness.

Limitations
The authors modestly note that the dataset was derived from a single batch under one controlled storage condition. Therefore, the reported prediction levels are internal estimates, and external validation across different cultivars, seasons, and commercial handling conditions is required before practical deployment. The study does not claim immediate commercial readiness but rather establishes a validated protocol and proof of concept for simplified NIR sensing systems.

More like this