光谱技术如何无损判定水果成熟？

How Can Spectral Technology Non-Destructively Determine Fruit Ripeness?

想知道牛油果何时入口最佳，榴莲是否熟透？水果成熟度检测一直是农业领域的重要课题，本篇将介绍光谱和高光谱成像技术如何为这一课题提供了创新解决方案，通过无损检测实现精准判断。

水果成熟过程中，其内部化学成分（如叶绿素、类胡萝卜素、糖分、酸度等）会发生规律性变化，这些物质对特定波长的光具有吸收和反射特性。研究表明，可见光波段主要反映色素变化，而近红外区域则与水分、糖分等内部成分密切相关。

在实际应用中，不同水果种类因其生理特性差异，需要采用特定的特征波长和不同的算法模型。

Want to know when an avocado is at its peak or if a durian is perfectly ripe? Fruit ripeness detection has always been a critical topic in agriculture. This article explores how spectral and hyperspectral imaging technologies provide innovative solutions for this challenge, enabling precise judgment through non-destructive testing.

During fruit ripening, internal chemical components (such as chlorophyll, carotenoids, sugars, acidity, etc.) undergo regular changes. These substances exhibit unique absorption and reflection characteristics for specific wavelengths of light. Research shows that the visible light spectrum primarily reflects pigment changes, while the near-infrared region is closely related to internal components like moisture and sugar content.

In practical applications, different fruit types require specific characteristic wavelengths and distinct algorithm models due to variations in their physiological properties.

小果的成熟度分析 / How Can Spectral Technology Non-Destructively Determine Fruit Ripeness?

一项甜橙研究采用400-1000nm波段的可见/近红外光谱，结合偏最小二乘法(PLS)，成功预测了可溶性固形物、可滴定酸和维生素C含量，为成熟期预测提供了量化依据。

香蕉成熟度检测常利用高光谱成像技术在400-1000nm范围内采集数据。一个研究团队通过主成分分析(PCA)结合极限学习机(ELM)建立的模型，对可溶性固形物和硬度的预测相关系数R²分别达到0.92和0.94。

针对牛油果的研究发现，其成熟度判断主要依赖于800nm以上的近红外信息，而520~650nm的可见光范围则有助于区分未成熟与成熟果实。研究人员开发的高光谱卷积神经网络(HS-CNN)模型，在牛油果成熟度分类中准确率超过90%。

A study on sweet oranges utilized visible/near-infrared spectroscopy in the 400–1000 nm range, combined with partial least squares (PLS), to successfully predict soluble solids, titratable acidity, and vitamin C content, providing a quantitative basis for ripening stage prediction.

For banana ripeness detection, hyperspectral imaging technology is often employed to collect data within the 400–1000 nm range. One research team developed a model using principal component analysis (PCA) combined with an extreme learning machine (ELM), achieving prediction correlation coefficients (R²) of 0.92 and 0.94 for soluble solids and firmness, respectively.

Research on avocados found that ripeness determination primarily relies on near-infrared information above 800 nm, while the visible light range of 520–650 nm helps distinguish unripe from ripe fruit. A hyperspectral convolutional neural network (HS-CNN) model developed by researchers achieved over 90% accuracy in avocado ripeness classification.

高光谱数据对牛油果的成熟度分类的决策影响：牛油果的空间维、光谱维图像 / The impact of the input on the decision of the class for an avocado

皮厚且坚硬的水果，如何检测？ / How to Detect Ripeness in Thick-Skinned, Hard Fruits?

对于西瓜、哈密瓜、榴莲等皮厚且坚硬的水果，成熟度检测面临很大的挑战。

一项西瓜研究使用了近红外光谱（NIRS）技术，涉及908~1676nm和950~1650nm光谱范围，检测了249个完整西瓜（152个浅绿条纹果皮，97个深绿纯色果皮）。利用偏最小二乘判别分析（PLS-DA），构建可溶性固形物含量（SSC）的定量模型。结果显示，浅绿条纹和深绿纯色西瓜的正确分类率分别为66.4%和82.2%，针对不同类型西瓜分别建立模型能获得更好结果。

哈密瓜与西瓜类似，研究显示，其可溶性固形物含量与特定波长反射率存在强相关性。通过优化选择的特征波长建立的简化模型，既保持了预测精度，又提高了检测速度。

榴莲作为巨大挑战性的厚皮水果之一，其成熟度检测一直依赖经验判断或破坏性方法。榴莲成熟度检测常依赖经验判断或破坏性方法。一项研究采用1100~2500nm光谱范围，使用果皮和茎的光谱信息对果肉干物质，进行间接预测成熟度。

研究发现，在将榴莲分为未成熟、早成熟和成熟类别的过程中，外皮模型更优；预测干物质含量方面，果皮模型表现更好。研究人员发现，尽管与参考果肉模型的精度相比，准确度相对较低，但在选定波长下，组合分析外皮和茎干光谱数据可提供较高分类精度。

A watermelon study employed near-infrared spectroscopy (NIRS) technology, covering spectral ranges of 908–1676 nm and 950–1650 nm, to examine 249 intact watermelons (152 with light green striped rinds and 97 with dark green solid rinds). Using partial least squares discriminant analysis (PLS-DA), a quantitative model for soluble solids content (SSC) was constructed. Results showed correct classification rates of 66.4% for light green striped watermelons and 82.2% for dark green solid ones, indicating that separate models for different types yield better outcomes.

Similar to watermelons, cantaloupe studies revealed strong correlations between soluble solids content and reflectance at specific wavelengths. Simplified models built with optimized characteristic wavelengths maintained prediction accuracy while improving detection speed.

Durian, one of the most challenging thick-skinned fruits, has traditionally relied on experiential judgment or destructive methods for ripeness assessment. A study used the 1100–2500 nm spectral range, leveraging rind and stem spectral data to indirectly predict pulp dry matter content as an indicator of ripeness.

The study found that for classifying durians into unripe, early ripe, and ripe categories, the rind model performed better. In predicting dry matter content, the rind model also showed superior performance. Researchers noted that although the accuracy was relatively lower compared to reference pulp models, combining rind and stem spectral data at selected wavelengths could achieve higher classification precision.

西瓜研究 / Watermelon Study：

西瓜的平均近红外光谱 / Average near-infrared spectra of watermelon

使用LVF仪器预测完整条纹浅绿色和实心深绿色外皮西瓜中可溶性固形物含量(%)的最佳方程的校准统计量 /

Calibration statistics of the optimal equation for predicting soluble solids content (%) in intact striped light-green and solid dark-green rind watermelons using LVF instrumentation

榴莲研究 / Durian Study：

(a)果柄和(b)果皮被放置在样品架中的情况。通过旋转旋钮可水平或垂直移动样品，如指针所示，使其到达检测焦点位置。

Photographs showing (a) the stem and (b) the rind placed in the sample holder. Knob rotations are used  to move the samples horizontally and vertically to the focal position for irradiation as indicated by the needle.

基于近红外光谱的不同成熟阶段均值光谱变化：(a) 果肉；(b) 果皮；(c) 果柄

Variationwithrespect to maturation stages of mean spectra using near-infrared spectroscopy of: (a) pulp;  (b) rind; and (c) stem.

讨论和结语 / Disscusion & Conclusion

实现稳健的校准模型是当前研究的重点。模型的低稳健性会阻碍其在跨环境（从实验室到现场）、跨样本（不同品种/年份）以及跨设备间的推广应用。自然界的复杂性和大量变异是主要挑战。可构建覆盖不同年份、果园和品种的多样化样品数据库增强模型适应性。

模型泛化能力研究仍显不足，实际生产中的环境条件波动也会进一步考验模型普适性。

作为高光谱成像系统硬件的提供商，我们致力于为高校研究所和解决方案集成商提供高性能、可靠的光学成像平台。我们期待与研究机构、系统集成商合作，共同开发面向特定场景的成熟度检测解决方案，推动这项技术从实验室走向田间和生产线。

Developing robust calibration models remains a key focus of current research. Low model robustness hinders their application across environments (from lab to field), samples (different varieties/years), and devices. The complexity and vast variability in nature pose major challenges. Building diverse sample databases covering multiple years, orchards, and varieties can enhance model adaptability.

Research on model generalization capability is still insufficient, and fluctuating environmental conditions in real-world production further test model universality.

As a provider of hyperspectral imaging system hardware, we are committed to delivering high-performance, reliable optical imaging platforms to academic institutions and solution integrators. We look forward to collaborating with research organizations and system integrators to develop ripeness detection solutions tailored to specific scenarios, advancing this technology from the lab to fields and production lines.

案例来源 / Source：

1. Varga LA, Makowski J, Zell A. Measuring the ripeness of fruit with hyperspectral imaging and deep learning. 2021 International Joint Conference on Neural Networks (IJCNN). 2021:1-8.

2. Vega-Castellote M, Sánchez MT, Torres I, de la Haba MJ, Pérez-Marín D. Assessment of watermelon maturity using portable new generation NIR spectrophotometers. Scientia Horticulturae. 2022;304:111328. 

3. Somton W, Pathaveerat S, Terdwongworakul A. Application of near infrared spectroscopy for indirect evaluation of “Monthong" durian maturity. International Journal of Food Properties. 2015;18(6):1155-1168. 

4. Liu J, Meng H. Research on the maturity detection method of Korla pears based on hyperspectral technology. Agriculture. 2024;14:1257.