{
  "id": 8318447,
  "title": "One-dimensional CNNs for Near-Infrared Prediction of Protein and Moisture in Cereal Grains: The Effects of Architecture and Input Preparation",
  "url": "https://urgent.news/2026/09/18/one-dimensional-cnns-for-near-infrared-prediction-of-protein-and",
  "topic": "science",
  "section": "Science",
  "published": "2026-09-18T00:00:00.000Z",
  "source": {
    "name": "bioRxiv",
    "slug": "biorxiv",
    "url": "https://www.biorxiv.org/content/10.64898/2026.09.12.751160v1?rss=1"
  },
  "original_language": "en",
  "account": "Near-infrared (NIR) spectroscopy is a popular technique for quickly and non-destructively determining key properties like protein content and moisture levels in cereal grains. However, comparing different models can be tricky when they don't process their data in the same way. To investigate this, researchers compared a simple one-dimensional convolutional neural network (CNN) to other models such as partial least squares regression (PLSR), support vector regression (SVR), and XGBoost across six different datasets containing protein and moisture data from cereal grains (with sizes ranging from 500 to 5046 samples). They tested three input scenarios: the raw spectra as provided, data that had been preprocessed for optimal performance, and data that had been preprocessed and had certain wavelengths selectively removed. The results showed that a CNN with just one filter had enough capability to perform well. When given the same raw data, the randomly searched CNN (CNN-RS) outperformed the compact CNN (CNN-Baseline) on five out of six datasets, while the PLSR model only performed best on the smallest protein dataset. Applying preprocessing techniques didn't generally improve CNN accuracy on the larger datasets, but did help a bit on the smaller ones, and shortened training time by up to 49%. Ultimately, the advantages of CNNs seem to come more from their robustness to raw input data than their absolute accuracy. When each model was given its own preprocessing and wavelength selection, SVR became the most accurate on four of the six datasets, and XGBoost performed nearly as well as the best model on four of the six datasets, indicating that the main benefit of CNNs may be their resilience rather than their precision. The effectiveness of a calibration model for NIR analysis of cereal grains seems to hinge more on the amount of data available and the preprocessing steps than the specific network architecture used.",
  "summary": "Near-infrared (NIR) spectroscopy is widely used for the rapid, non-destructive determination of constituents such as protein and moisture in cereal grains, but model comparisons in this field are often confounded by differences in the inputs received by each model. We benchmark a compact one-dimensional CNN derived by one-factor-at-a-time ablation (CNN-Baseline) and a randomly searched CNN…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}