當(dāng)我們描述一杯咖啡帶有花香、堅(jiān)果或焦糖風(fēng)味時(shí)，是否想過(guò)，這份獨(dú)特的風(fēng)味能否在不依賴杯測(cè)的情況下，于烘焙前就能被預(yù)測(cè)？

高光譜成像技術(shù)正在將這種構(gòu)想變?yōu)榭赡堋８吖庾V能無(wú)損地掃描咖啡生豆，獲得其完整的光譜數(shù)據(jù)，研發(fā)人員從光譜中找出與特定風(fēng)味物質(zhì)（如糖分、caffeine）之間的定量關(guān)系，以此為基礎(chǔ)，構(gòu)建出可靠的預(yù)測(cè)模型。通過(guò)這些模型，我們得以科學(xué)地預(yù)見(jiàn)咖啡豆的風(fēng)味輪廓。

在一項(xiàng)研究中，科研人員將單顆咖啡豆（生豆）放置在黑色樣品臺(tái)上，利用高光譜相機(jī)對(duì)其進(jìn)行掃描。高光譜相機(jī)的覆蓋短波紅外區(qū)域（900-2500nm），咖啡豆內(nèi)部的主要化學(xué)成分，如糖分、生物堿、油脂等，在此波段下會(huì)呈現(xiàn)出獨(dú)特的“光譜指紋"。利用專業(yè)的軟件分析和建立模型，定量推測(cè)豆子中各種物質(zhì)的含量。

研究人員掃描了多個(gè)產(chǎn)區(qū)的數(shù)百顆樣本，以檢測(cè)蔗糖（甜味）、caffeine（苦味）和葫蘆巴堿（烘焙香氣）這三種關(guān)鍵風(fēng)味前體。

通過(guò)結(jié)合高光譜數(shù)據(jù)與液相色譜質(zhì)譜法的精密測(cè)量值，應(yīng)用PLSR算法，研究團(tuán)隊(duì)構(gòu)建了上述成分的定量預(yù)測(cè)模型。交叉驗(yàn)證結(jié)果表明，模型對(duì)caffeine和葫蘆巴堿的預(yù)測(cè)精度很高（R2 > 0.8），對(duì)蔗糖的預(yù)測(cè)也可用于初步篩選。

利用高光譜成像的特性，團(tuán)隊(duì)還生成了這些化學(xué)成分在豆子內(nèi)部的空間分布圖，直觀顯示了它們的不均勻分布，這為了解咖啡豆的生理結(jié)構(gòu)提供了新視角。

通過(guò)高光譜成像（HSI）獲取的純品參比物質(zhì)（caffeine、蔗糖和葫蘆巴堿）以及磨碎生咖啡豆樣品的平均光譜，同時(shí)展示了1400nm單一光譜波段的吸光度圖像（右側(cè)）

該研究團(tuán)隊(duì)還轉(zhuǎn)向烘焙后咖啡豆，為應(yīng)對(duì)更復(fù)雜的風(fēng)味分析，采用了可同時(shí)預(yù)測(cè)多個(gè)響應(yīng)變量的PLS2算法。研究人員不僅建立了針對(duì)醛類、吡嗪類等化學(xué)族群的預(yù)測(cè)模型，更進(jìn)一步將化合物按其感官特征（如堅(jiān)果香、甜香）分組，建立了針對(duì)整體風(fēng)味屬性的模型。結(jié)果表明，對(duì)醛類（甜香）和吡嗪類（烘烤香）等族群的預(yù)測(cè)效果尤為出色。

為了驗(yàn)證這一預(yù)測(cè)能力的實(shí)際價(jià)值，研究人員進(jìn)行了一項(xiàng)實(shí)驗(yàn)：他們利用建立好的模型，對(duì)一批阿拉伯比卡咖啡豆進(jìn)行掃描，并根據(jù)模型預(yù)測(cè)的吡嗪含量和堅(jiān)果香氣強(qiáng)度，手動(dòng)篩選出預(yù)測(cè)值max和min的10%的豆子，分別組成新的批次。

對(duì)這兩個(gè)批次豆子的化學(xué)分析證實(shí)，分選效果極其顯著。被預(yù)測(cè)為“高吡嗪"的批次，其實(shí)際吡嗪類物質(zhì)含量顯著高于原始混合批次和“低吡嗪"批次。

這證明，高光譜成像技術(shù)結(jié)合預(yù)測(cè)模型，能夠有效識(shí)別咖啡豆內(nèi)部的特定風(fēng)味物質(zhì)差異。該研究為在產(chǎn)業(yè)線上實(shí)現(xiàn)非破壞性的精準(zhǔn)風(fēng)味分選提供了新思路，展現(xiàn)了其用于生產(chǎn)風(fēng)味定制化咖啡產(chǎn)品的技術(shù)潛力，但其大規(guī)模穩(wěn)定應(yīng)用的效能仍有待進(jìn)一步驗(yàn)證。

按A.預(yù)測(cè)吡嗪類化合物含量或B.分析預(yù)測(cè) 堅(jiān)果味（粗體標(biāo)注）的前面10%（高含量組，H）或后10%（低含量組，L）對(duì)咖啡豆進(jìn)行分選后，分選試驗(yàn)對(duì) 4 組揮發(fā)性化合物（吡嗪類、醛類、酮類和雜環(huán)含氮化合物）相對(duì)豐度及分析預(yù)測(cè)的堅(jiān)果味、果香味、酸味和烘焙味的影響。

從咖啡生豆的檢測(cè)，到熟豆后天風(fēng)味的預(yù)測(cè)，高光譜成像技術(shù)為我們提供了一條貫穿咖啡品質(zhì)管控全程的強(qiáng)大紐帶。它讓“看豆識(shí)風(fēng)味"成為可能，將咖啡的品質(zhì)控制從傳統(tǒng)依賴經(jīng)驗(yàn)的“批量評(píng)估"，推向了一個(gè)數(shù)字化、可視化的“單顆精準(zhǔn)管理"新時(shí)代。

案例來(lái)源：

Caporaso, N., Whitworth, M. B., Grebby, S., & Fisk, I. D. (2018). Non-destructive analysis of sucrose, caffeine and trigonelline on single green coffee beans by hyperspectral imaging. Food Research International, 106, 193–203.

Caporaso, N., Whitworth, M. B., & Fisk, I. D. (2022). Prediction of coffee aroma from single roasted coffee beans by hyperspectral imaging. Food Chemistry, 371, 131159.

Can Hyperspectral Imaging Detect the Flavor of Coffee Beans?

When describing a cup of coffee as having floral, nutty, or caramel notes, have you ever wondered whether these unique flavors could be predicted before roasting—without relying on cupping?

Hyperspectral imaging technology is turning this concept into a reality. By scanning green coffee beans non-destructively, hyperspectral systems capture complete spectral data. Researchers identify quantitative relationships between the spectral signatures and specific flavor compounds—such as sugars and caffeine—and use these to build reliable predictive models. Through these models, the flavor profile of coffee beans can be scientifically anticipated.

In one study, researchers placed individual green coffee beans on a black sample stage and scanned them using a hyperspectral camera. The camera covered the short-wave infrared range (900–2500 nm), where key chemical components inside the beans—such as sugars, alkaloids, and oils—exhibit distinct "spectral fingerprints." Using specialized software for analysis and modeling, the content of various compounds in the beans was quantitatively estimated.

Hundreds of samples from multiple global growing regions were scanned to detect three key flavor precursors: sucrose (sweetness), caffeine (bitterness), and trigonelline (roasty aroma).

By combining hyperspectral data with precise measurements from liquid chromatography–mass spectrometry, the research team applied PLSR (Partial Least Squares Regression) to develop quantitative prediction models for these components. Cross-validation results showed high prediction accuracy for caffeine and trigonelline (R2 > 0.8), and sucrose predictions were suitable for preliminary screening.

Leveraging the capabilities of hyperspectral imaging, the team also generated spatial distribution maps of these chemical compounds inside the beans. These visualizations clearly revealed their uneven distribution, offering new insights into the physiological structure of coffee beans.

The average spectra of pure reference substances (caffeine, sucrose, and trigonelline) and ground green coffee bean samples obtained through hyperspectral imaging (HSI), along with the absorbance image at a single spectral band of 1400 nm (shown on the right).

The research team also examined roasted coffee beans. To address more complex flavor analysis, they employed the PLS2 algorithm, which can predict multiple response variables simultaneously. The researchers not only built prediction models for chemical groups such as aldehydes and pyrazines, but also grouped compounds by sensory attributes—such as nutty aroma and sweet aroma—to develop models targeting overall flavor characteristics. Results indicated particularly strong predictive performance for groups like aldehydes (sweet aroma) and pyrazines (roasty aroma).

To test the practical value of this predictive capability, the researchers conducted an experiment: using the established model, they scanned a batch of Arabica coffee beans. Based on the predicted pyrazine content and nutty aroma intensity, they manually selected the top 10% and bottom 10% of beans to form two new batches.

Chemical analysis of these two batches confirmed highly significant sorting results. The batch predicted as "high pyrazine" showed substantially higher actual pyrazine content compared to the original mixed batch and the "low pyrazine" batch.

This demonstrates that hyperspectral imaging combined with predictive models can effectively identify differences in specific flavor compounds within coffee beans. The study offers a new approach for non-destructive, precise flavor sorting on industrial production lines, highlighting the technology’s potential for producing customized coffee products. However, the efficiency and stability of large-scale application still require further validation.

After sorting coffee beans into either the top 10% (high-content group, H) or the bottom 10% (low-content group, L) based on A. predicted pyrazine content or B. predicted "nutty flavor" (indicated in bold), the sorting experiment’s impact on the relative abundance of four groups of volatile compounds (pyrazines, aldehydes, ketones, and nitrogen-containing heterocycles) and the predicted intensities of nutty flavor, fruity flavor, acidity, and roasty flavor.

From detecting green coffee beans to predicting the developed flavors of roasted beans, hyperspectral imaging provides a powerful link throughout the entire coffee quality control process. It makes it possible to "see flavor in the bean," shifting coffee quality control from traditional, experience-based "batch assessment" toward a new era of digital, visual, and "single-bean precision management."

Sources：

Caporaso, N., Whitworth, M. B., Grebby, S., & Fisk, I. D. (2018). Non-destructive analysis of sucrose, caffeine and trigonelline on single green coffee beans by hyperspectral imaging. Food Research International, 106, 193–203.

Caporaso, N., Whitworth, M. B., & Fisk, I. D. (2022). Prediction of coffee aroma from single roasted coffee beans by hyperspectral imaging. Food Chemistry, 371, 131159.