When "Mild Spicy" in Sichuan Equals "Extremely Spicy" in Guangdong... Hyperspectral Camera Measures Chili Heat Levels

每個人對辣度的接受程度都不一樣，火鍋底料的辣度如何科學量化？本次實驗利用高光譜相機，對6種不同辣度的火鍋底料進行測試，探索光譜數(shù)據(jù)與辣度的關(guān)聯(lián)性。

People's tolerance for spiciness varies widely, but how can the heat level of hot pot base be scientifically quantified? This experiment utilized a hyperspectral camera to test six hot pot bases with different spiciness levels, exploring the correlation between spectral data and chili heat intensity.

「樣品介紹 / Samples」

測試6種不同辣度的火鍋底料，辣度分別為：12°、36°、45°、52°、65°、75°

Six hot pot base samples with varying heat levels were tested: 12°, 36°, 45°, 52°, 65°, and 75°.

「數(shù)據(jù)采集 / Data Acquisition」

高光譜相機：覆蓋400~1700nm波段（可見光+短波紅外）

成像方式：線性推掃，確保數(shù)據(jù)精準

光源與環(huán)境：鹵素燈均勻照明，暗室環(huán)境減少干擾

樣品擺放：水平位移臺固定，保證成像穩(wěn)定

Hyperspectral Camera: Covered 400–1700 nm (visible light + short-wave infrared).

Imaging Method: Linear push-broom scanning for precise data capture.

Lighting & Environment: Halogen lamp for uniform illumination, darkroom to minimize interference.

Sample Setup: Fixed on a horizontal displacement platform for stable imaging.

400-1000nm

900-1700nm

「分析方法 / Analysis Method」

高光譜成像不僅能拍出照片，還能記錄每個像素點的光譜“指紋"。 

實驗過程中，首先使用400-1000nm可見近紅外和900-1700nm短波紅外兩臺高光譜相機采集6種火鍋底料樣品的光譜數(shù)據(jù)。

在數(shù)據(jù)預處理階段，通過專業(yè)的高光譜分析軟件對原始數(shù)據(jù)進行降噪處理和反射率計算，同時消除背景光譜干擾，確保獲得純凈的目標物體光譜信息，這一過程通常在數(shù)據(jù)采集時同步完成。

隨后從處理后的高光譜數(shù)據(jù)中提取關(guān)鍵特征，包括光譜反射率、吸收峰位置及光譜形態(tài)特征等，并運用主成分分析等降維方法篩選出具有代表性的特征參數(shù)。

在分類識別環(huán)節(jié)，利用不同物質(zhì)對特定波段反射率的差異特性，分別采用監(jiān)督學習和無監(jiān)督學習兩種方法：前者通過標記數(shù)據(jù)集訓練光譜角制圖或卷積神經(jīng)網(wǎng)絡(luò)等分類模型，后者則運用K均值或?qū)哟尉垲惖人惴▽崿F(xiàn)數(shù)據(jù)自動分類。

最終將分析結(jié)果以偽彩色圖像形式直觀呈現(xiàn)，展示不同物質(zhì)的空間分布情況，并基于光譜特征開展定量和定性分析，計算得出各類物質(zhì)的濃度或類別參數(shù)。

Hyperspectral imaging not only captures photos but also records the spectral "fingerprint" of each pixel.

During the experiment, two hyperspectral cameras (400–1000 nm visible-NIR and 900–1700 nm SWIR) were used to collect spectral data from the six samples.

In the preprocessing stage, raw data underwent noise reduction and reflectance calibration via specialized software, while background interference was eliminated to ensure clean spectral data. This process was synchronized with data acquisition.

Key features were then extracted from the processed data, including spectral reflectance, absorption peak positions, and spectral shape characteristics. Dimensionality reduction methods like PCA were applied to identify the most representative parameters.

For classification, both supervised and unsupervised learning were employed:

Supervised methods (e.g., spectral angle mapper or CNN) used labeled datasets to train models.

Unsupervised methods (e.g., K-means or hierarchical clustering) automated data grouping based on reflectance differences in specific bands.

Results were visualized as pseudo-color images to display spatial distributions of materials, followed by quantitative/qualitative analysis to calculate concentrations or categories.

「光譜曲線 / Spectral Curves」

在400-1700nm波長范圍內(nèi)，六種不同辣度的火鍋底料樣本在a面和b面的反射率曲線呈現(xiàn)出相似的光譜波形，但反射率數(shù)值隨辣度變化而存在顯著差異。具體表現(xiàn)為辣度越高，反射率越低，這一趨勢在a面和b面均保持一致。

值得注意的是，在b面的860-930nm波段范圍內(nèi)，反射率曲線對辣度的區(qū)分效果尤為明顯，能夠更清晰地反映辣度差異。

Within 400–1700 nm, reflectance curves of the six samples (A-side and B-side) showed similar waveforms but significant reflectance variations correlated with spiciness. Higher heat levels consistently exhibited lower reflectance on both sides.

Notably, the 860–930 nm range on the B-side provided the clearest distinction between heat levels.

a面反射率(400-1000nm)

a面反射率(900-1700nm)

b面反射率(400-1000nm)

「建立CNN模型 / CNN Modeling」

為了進一步分析辣度分類的可行性，研究采用卷積神經(jīng)網(wǎng)絡(luò)（CNN）對高光譜數(shù)據(jù)進行建模。

To further assess classification feasibility, a CNN model was applied to hyperspectral data.

建立CNN模型（400-1000nm  a/b面） / CNN Model (400–1000 nm, A/B-sides)

在400-1000nm波段的a面數(shù)據(jù)分類中，模型整體準確率介于75%-85%之間，其中辣度45°和75°的分類效果較好，而辣度12°和52°由于數(shù)據(jù)采集時受容器遮擋影響，部分區(qū)域出現(xiàn)誤判。此外，辣度36°因樣品表面凹陷導致數(shù)據(jù)質(zhì)量下降，而辣度65°的部分區(qū)域被錯誤歸類為75°。

相比之下，b面的分類表現(xiàn)更為穩(wěn)定，整體準確率約為85%，僅辣度45°的少量區(qū)域被誤判為36°。

A-side: Overall accuracy ranged 75%–85%. Samples at 45° and 75° were classified best, while 12° and 52° suffered partial misclassification due to container obstruction during imaging. The 36° sample had uneven surfaces, and 65° was occasionally mislabeled as 75°.

B-side: Performance was more stable (~85% accuracy), with only minor misclassification (45° vs. 36°).

a面結(jié)果 (400-1000nm)

a面結(jié)果 (400-1000nm)

建立CNN模型（900-1700nm  a/b面） / CNN Model (900–1700 nm, A/B-sides)

在900-1700nm波段的分析中，a面數(shù)據(jù)的分類準確率在70%-80%之間，其中辣度12°和36°因表面凹凸不平或凹陷導致數(shù)據(jù)質(zhì)量較差，誤判率較高，而辣度45°、52°、65°和75°的分類效果較好。

相比之下，b面數(shù)據(jù)由于表面更平滑，且無干辣椒等固體遮擋，分類表現(xiàn)顯著優(yōu)于a面，整體準確率超過90%，僅有少量區(qū)域出現(xiàn)誤判。

這一結(jié)果表明，900-1700nm波段可能更適合用于火鍋底料辣度的精準檢測，尤其是結(jié)合b面數(shù)據(jù)時，分類效果更佳。

A-side: Accuracy was 70%–80%. Samples at 12° and 36° showed higher misclassification due to surface irregularities, while 45°–75° performed better.

B-side: Superior accuracy (>90%) was achieved thanks to smoother surfaces and absence of solid obstructions (e.g., dried chilies).

These results suggest that 900–1700 nm SWIR, especially with B-side data, is more suitable for precise heat-level detection.

a面結(jié)果 (900-1700nm)

b面結(jié)果 (900-1700nm)

「總結(jié) / Conclusion」

基于高光譜視覺技術(shù)的研究表明，通過對六種不同辣度的火鍋底料樣品進行高光譜數(shù)據(jù)采集，并經(jīng)過數(shù)據(jù)預處理和算法分析，能夠有效區(qū)分樣品的辣度等級。

實驗數(shù)據(jù)顯示，雖然樣品a面和b面的光譜曲線均能反映辣度變化，但b面的區(qū)分效果更為顯著。在光譜波段選擇方面，相比400-1000nm的可見近紅外譜段，900-1700nm的短波紅外譜段展現(xiàn)出更高的識別準確率和檢測精度。

為進一步提升研究結(jié)果的可靠性，后續(xù)工作將重點擴大樣本數(shù)據(jù)量，通過增加樣本多樣性來持續(xù)優(yōu)化識別準確率。

Hyperspectral imaging effectively differentiated the six heat levels of hot pot base samples after data preprocessing and algorithmic analysis.

While both A-side and B-side spectral curves reflected spiciness trends, the B-side provided clearer distinctions. Compared to 400–1000 nm visible-NIR, the 900–1700 nm SWIR band demonstrated higher accuracy and precision.

To enhance reliability, future work will expand sample diversity and dataset size for further optimization.