A new deep-learning framework that was recently developed improves the accuracy of near-infrared spectroscopy for non-destructively measuring internal quality traits.
To measure the internal quality of fruits and grains accurately, analytical frameworks need to be accurate. A recent study explored this topic to see how machine learning (ML) can play a role in improving accuracy. Researchers at Henan University of Technology tested their newly developed ML framework that was designed to improve the accuracy of
In their study, the researchers tested how their model, known as multiscale deep ensemble regression (MDER), performed against two agricultural data sets. The two data sets the researchers looked at in the study were mangoes measured for dry matter content across multiple growing seasons, and corn kernels measured for protein content across multiple growing regions.1
In both cases, MDER produced more accurate predictions than is typical of conventional chemometric approaches, achieving an R² value of 0.9352 for mango dry matter and 0.9046 for corn protein on test data.1 These are figures that indicate a strong correlation between the model's predictions and actual laboratory measurements.
Increasingly, NIR spectroscopy being used in agriculture and food processing because it can assess internal quality attributes, such as sugar content, moisture, or protein, without cutting open or destroying the sample.1–4 That makes it valuable for sorting and grading produce on packing lines or evaluating grain shipments quickly.1
However, NIR spectroscopy is not a perfect technique for this space. One of the major, ongoing challenges is how the molecular absorption bands overlap, which is known as spectral collinearity.1 This is especially the case when it is evaluating crops grown in different seasons, regions, or storage conditions.1–4
For buyers, processors, and quality-control operations relying on NIR instruments, this has meant that models calibrated on one batch or season often perform poorly when applied to a different one.1
The first thing to note about MDER is its design. It is comprised of five one-dimensional
The outputs from these branches are then combined by a "meta-learner."1 The meta-learner is a smaller neural network equipped with attention mechanisms and residual connections that re-weights the contributions from each branch based on how predictive they are.1 According to the study, this step suppresses spectral signals that are collinear but chemically uninformative, while amplifying the signals that actually correspond to the compositional trait being measured.1
The authors believe that MDER potentially can be used to handle the mismatch between chemically diverse agricultural samples and their spectrally similar NIR readings. As the study indicated here, MDER could be applied to other crop types and different compositional targets.1
Another future study could also test MDER in live production settings. The researchers only validated the MDER model on curated research data sets.1 As a result, the model needs to be tested to confirm performance gains and how they carry over to real-world sorting lines.1 This is important because variables such as lighting, sample presentation, and instrument variability introduce additional sources of error.1
What this study does show definitively, though, is that NIR spectroscopy will continue to be used in food analysis applications.1,2
