The research team, working with 49 cassava germplasm accessions from the National Cassava Germplasm Repository in Danzhou, Hainan Province, China, set out to build a complete pipeline for grading storage roots by amylose content. In April 2025, they harvested roots at least 25 centimeters long with mid-section diameters of 5.5 centimeters or more, sliced them into uniform 3-centimeter sections with the peel intact, and imaged the exposed cut surfaces. Each accession contributed six root-derived samples, one from each of three anatomical regions across two roots, yielding 294 samples in total. After imaging, the tissue was chopped, homogenized, flash-frozen in liquid nitrogen, and stored at minus 40 degrees Celsius for reference chemical analysis.

The imaging system itself was a serious piece of laboratory hardware. A slit width of 12 micrometers and a spectral resolution of 2.1 nanometers produced 300 spectral bands spanning the visible and near-infrared range, while eight 150-watt halogen lamps arranged in two symmetric rows bathed the samples in uniform light. A motorized translation stage carried the samples past the sensor at 2 millimeters per second, and each resulting data cube weighed in at 1.8 gigabytes. Raw images were radiometrically calibrated against dark references and a standard white panel to strip out illumination noise before any analysis began.

Reference amylose values came from the iodine–dimethyl sulfoxide spectrophotometric method, with absorbance measured at 600 nanometers against a calibration curve. Using expert-consulted thresholds, samples were sorted into three classes: below 10.21 percent amylose as low, 10.21 to below 20.46 percent as medium, and 20.46 percent or above as high. These labels became the ground truth for supervised classification. Crucially, the team also had to decide what part of each sliced root the camera should actually look at, because peel and flesh differ dramatically in both chemistry and optics.

That segmentation problem was solved in two stages. First, pseudo-color images built from three spectral bands at 639, 549, and 470 nanometers were binarized with Otsu’s adaptive thresholding, and the six largest connected components were retained to remove background clutter. Then a SegFormer semantic segmentation network, trained on manually annotated peel regions, carved out the peel with impressive precision, achieving an Intersection over Union of 0.8927, a precision of 0.9187, and a Dice coefficient of 0.9403. Subtracting the peel mask from the whole-section mask left a flesh-only region, from which a single mean reflectance spectrum was extracted per sample. The mean spectra confirmed that peel reflectance runs consistently higher than flesh across the entire range, with the largest gaps in the visible region, validating the decision to separate the tissues before modeling.

With spectra in hand, the researchers benchmarked four classifiers—logistic regression, support vector machine, random forest, and XGBoost—against raw spectra and four preprocessing schemes: Savitzky–Golay smoothing, standard normal variate transformation, and first- and second-derivative transforms. The verdict was clear. XGBoost on raw or Savitzky–Golay-smoothed spectra delivered the best result, 83.99 percent accuracy with a standard deviation of just 0.80 percent across five-fold cross-validation, where folds were grouped by accession so that no variety leaked between training and test sets. Aggressive transformations generally hurt: second-derivative preprocessing dragged support vector machines down to 70.45 percent, suggesting that amplifying local spectral variation also amplifies noise.

The next step tackled the black-box problem. Nonlinear models may predict well, but breeders need to know which wavelengths carry the signal. Enter Shapley Additive Explanations, a game-theory-based interpretability framework that quantifies each spectral band’s contribution to every prediction. SHAP analysis revealed that amylose discrimination rests on a compact set of bands rather than the full 300-dimensional spectrum, with the ten most consistently important wavelengths clustered between roughly 719 and 1028 nanometers, including 789.46, 798.06, 719.06, 748.83, and 1001.23 nanometers. These regions overlap with bands reported for starch-related traits in maize and wheat, lending physiological plausibility to the selection. Tree-based models showed the sharpest concentration of importance, with XGBoost the most sparse of all, consistent with its boosting-driven feature selection.

The study’s most ambitious move followed from that insight: instead of reconstructing the entire hyperspectral cube from RGB images, the team trained a network to recover only the SHAP-selected, task-relevant bands. RGB images were simulated from the hyperspectral data using the CIE 1931 standard observer and D65 illuminant, then converted to sRGB. Building on the Multi-stage Spectral-wise Transformer++ architecture, the authors added two custom modules. The Masked Grouped Spectral Self-Attention module shuffles and groups spectral channels to capture cross-group correlations efficiently, while a foreground mask prevents background pixels from contaminating the attention calculation. The Spectral Feature-wise Linear Modulation module injects wavelength-dependent positional encodings as conditioning signals, predicting per-band scaling and bias terms that adaptively tune feature responses. Together, these modules cut the mean relative absolute error from 0.0731 for the baseline MST++ to 0.0697, lifted the peak signal-to-noise ratio from 25.82 to 26.57 decibels, and outperformed HDNet and MIRNet baselines across five independent training runs.

The decisive test was whether spectra reconstructed from cheap RGB images could still grade amylose. Using measured reflectance at the selected bands, random forest achieved 77.69 percent accuracy and an F1-score of 77.75 percent. With reconstructed reflectance, XGBoost led at 71.75 percent accuracy and an F1-score of 71.57 percent, a drop of only about 4.7 percentage points from the measured-band result. Two control experiments sealed the argument. Classification from 37 conventional RGB texture and histogram features topped out at 56.04 percent accuracy, and reconstruction of randomly chosen bands reached only 55.03 percent with XGBoost. The gap—more than 16 percentage points in both comparisons—demonstrates that the performance genuinely stems from recovering amylose-relevant spectral information, not from incidental image cues or generic reconstruction quality.

The authors are candid about the limits. The RGB inputs were simulated under standardized color-rendering conditions rather than captured by real cameras, and practical deployment will introduce domain shifts from sensor response, illumination, exposure, and white balance. Robustness across harvesting years, geographic origins, and imaging devices remains untested, and the study did not exhaustively compare classical chemometric classifiers or alternative wavelength-selection strategies. Still, the framework marks a meaningful shift in how agricultural spectral sensing might be built. By letting interpretability analysis decide which wavelengths matter, then reconstructing only those bands and validating the result at the task level rather than by spectral similarity metrics alone, the approach offers a cost-aware template for high-throughput phenotyping. For a crop that anchors food security across the tropics, the prospect of grading roots with nothing more than a camera and a trained network is a compelling glimpse of where digital agriculture is heading.

Subject of Research: Hyperspectral imaging and RGB-based spectral reconstruction for grading cassava storage roots by amylose content

Article Title: Slice-based grading of cassava storage roots by amylose content using hyperspectral analysis and RGB-to-hyperspectral reconstruction

Article References: Chen, Y., Wei, Y., Wang, Y., Fu, N., Fang, J., Wang, B., Zhang, J., Wan, Z., Xue, M., Tian, T., Ye, J., & Feng, H. (2026). Slice-based grading of cassava storage roots by amylose content using hyperspectral analysis and RGB-to-hyperspectral reconstruction. Smart Agricultural Technology, 15, Article 102602. https://doi.org/10.1016/j.atech.2026.102602

Image Credits: AI Generated

DOI: Not provided

Keywords: cassava, amylose, hyperspectral imaging, RGB reconstruction, machine learning, XGBoost, SHAP, deep learning, food security, crop phenotyping, spectral band selection, agricultural technology