Key points
- The study collected 67 cassava varieties and tested whether near-infrared spectroscopy (NIRS) could predict flour quality across the 11,000-4,000 cm⁻¹ range
- The protein and moisture models performed well, with cross-validated R² values of 0.87 and 0.83 respectively
- The starch, amylose and amylopectin models reached R² of only 0.36, 0.34 and 0.19, so the authors restrict them to preliminary screening
- The fat model performed worst at R² 0.16 and was judged unsuitable for prediction
- The authors stress that the wide variation in performance between constituents shows external validation is needed
The question behind the study
Cassava is the sixth most important food crop globally, and its value rests on the starch it accumulates in its storage roots.
Quality evaluation is therefore essential to how the crop is used, yet rapid and efficient assessment methods remain underexplored.
Researchers in China set out to test how far near-infrared measurement — which is fast and does not destroy the sample — could stand in for laboratory analysis.
Method and results
Flour samples from the 67 varieties were modelled using partial least squares together with principal component analysis and internal cross-validation.
The results split cleanly into two groups: constituents the technique predicts well, and constituents it does not.
| Constituent | Cross-validated R² | RPD | Authors’ classification |
|---|---|---|---|
| Protein | 0.87 | 10.26 | suitable for preliminary quantitative analysis |
| Moisture | 0.83 | 4.36 | suitable for preliminary quantitative analysis |
| Starch | 0.36 | 2.49 | preliminary screening only |
| Amylose | 0.34 | 2.21 | preliminary screening only |
| Amylopectin | 0.19 | 2.36 | preliminary screening only |
| Fat | 0.16 | 1.08 | unsuitable for prediction |
Source: Guo et al., Foods vol. 15 no. 17 (2026)
The cross-validated R² indicates how much of the variation in the true values a model explains, while RPD is the ratio between the spread of the data and the error of the prediction.
The authors report that the starch, amylose and amylopectin models showed substantial overfitting, and that the fat model showed severe overfitting.
How to read this carefully
These measurements were taken on prepared cassava flour, not on fresh roots at a buying point.
Measuring the starch content of fresh roots at a Thai factory gate is a different context altogether — different sample condition, different moisture, and different established methods.
The study therefore answers a question about screening variety quality, not a question about measuring starch content to set a buying price.
The authors themselves describe the framework as exploratory, and say the marked variability in predictive performance across constituents makes external validation necessary to improve model robustness.
Why growers and buyers should care
Starch content is the variable that sets the value of a cassava root directly, so a fast, non-destructive way to measure it is something the whole chain wants.
What this study reports is that the technique is not yet ready for the starch constituents themselves, even though it works well for protein and moisture.
For variety selection work, fast preliminary screening still has value, even where the readings must be confirmed by a standard method.
Readers interested in quality standards and controlled parameters for tapioca starch can see our certifications page.
The research described here was verified through the PubMed database.
ที่มา / Sources:
- Guo H., Wang W., Liu C., Guo S., Ou W., Gao F., Li K., Zhang L., Ren G. “Establishment of a Near-Infrared Spectroscopy-Based Screening Framework for Key Quality Indicators of Cassava Varieties.” Foods vol. 15 no. 17, 4 September 2026 (verified via PubMed, PMID 42737382) - https://doi.org/10.3390/foods15173139
- PubMed, biomedical bibliographic database - https://pubmed.ncbi.nlm.nih.gov/42737382/