When Grapes Ripen: Case Studies of Hyperspectral Field Detection

在(zài)葡(pú)萄(táo)種植與(yǔ)葡萄酒(jiǔ)釀造領域(yù)，準確監測(cè)葡(pú)萄成熟度(dù)和糖度是決(jué)定(dìng)采收時(shí)機和(hé)最(zuì)終酒品(pǐn)質(zhì)量的關(guān)鍵(jiàn)。高光譜(pǔ)成像技(jì)術(shù)已(yǐ)被(bèi)證明(míng)能(néng)夠有(yǒu)效檢測葡(pú)萄(táo)的(dí)成熟度(dù)，糖(táng)度(°Brix)和花青素含量，作為高(gāo)光譜成像技(jì)術的專(zhuān)業提(tí)供商(shāng)，我們(mén)期待看到這(zhè)項技術在(zài)葡(pú)萄品質監測中的創新(xīn)應用。

目前(qián)大多數研(yán)究都局(jú)限(xiàn)於實驗室環境(jìng)或(huò)對單(dān)顆葡(pú)萄進行檢(jiǎn)測。今(jīn)天，我們將聚(jù)焦兩(liǎng)項突(tū)破性(xìng)研究，展示高光譜技術如何在(zài)田間(jiān)現(xiàn)場(cháng)實現葡萄(táo)品質的非破壞性監(jiān)測(cè)。

In viticulture and winemaking, accurately monitoring grape maturity and sugar content is crucial for determining optimal harvest timing and final wine quality. Hyperspectral imaging technology has proven effective in detecting grape maturity, sugar content (°Brix), and anthocyanin levels.As a professional provider of hyperspectral imaging solutions, we are excited to see innovative applications of this technology in grape quality monitoring.

Currently, most research has been limited to laboratory settings or single-berry measurements. Today, we highlight two groundbreaking studies demonstrating how hyperspectral technology enables non-destructive quality monitoring directly in the field.

『從實驗室走(zǒu)向(xiàng)田間(jiān) / From Lab to Vineyard 』

意大(dà)利團隊開發了一(yī)種基於(yú)可(kě)見光(guāng)-近紅外(wài)(400-1000nm)高光(guāng)譜(pǔ)相(xiāng)機(jī)的(dí)非破(pò)壞(huài)性方法，直接(jiē)在葡萄園中進(jìn)行(háng)13填(tián)的連續監測(cè)。他們使用偏(piān)最小二(èr)乘回(huí)歸(guī)(PLS)預測可溶性固(gù)形物含量(liáng)，獲得R²=0.77的(dí)預測(cè)精度(dù)(RMSECV=0.79°Brix)，並通過(guò)偏最小二乘判別分(fēn)析(PLS-DA)將(jiāng)葡(pú)萄按成熟度(dù)(以(yǐ)20°Brix為界(jiè))分(fēn)類(lèi)，準(zhǔn)確(què)率達86-91%。

An Italian research team developed a non-destructive method using visible-to-near-infrared (400-1000nm) hyperspectral cameras for continuous 13-day monitoring in vineyards. Using partial least squares regression (PLS), they achieved soluble solids content (SSC) predictions with R²=0.77 (RMSECV=0.79°Brix). Through partial least squares discriminant analysis (PLS-DA), they classified grape maturity (using 20°Brix as the threshold) with 86-91% accuracy.

圖(tú)1 / Figure 1
：

(a) RGB圖像來源於葡萄園(yuán)行掃描截麵(miàn)的高(gāo)光譜數據(jù)。 (b) 基於曼(màn)哈(hā)頓函(hán)數(shù)分(fēn)類生(shēng)成的(dí)ROI（紅(hóng)色(sè)區域）。

(a) The RGB image is derived from hyperspectral data captured during a vineyard row scan transect.

(b) The region of interest (ROI, marked in red) was generated through classification using the Manhattan distance function.

『邊(biān)走(zǒu)邊(biān)測
：硬(yìng)件與(yǔ)算(suàn)法(fǎ)的(dí)結(jié)合(hé) / “On the Go" Monitoring: Hardware and Algorithm Synergy 』

西班牙(yá)團(tuán)隊(duì)的(dí)研究(jiū)則更(gēng)進(jìn)一步(bù)，他(tā)們開發(fā)了一套創新的"行進間"高光(guāng)譜成(chéng)像係統(tǒng)，將可見光(guāng)-近紅外(wài)高(gāo)光譜(pǔ)相機（400-1000nm，光(guāng)譜分辨(biàn)率2.1nm的）安裝在(zài)以5公裏/小(xiǎo)時速度(dù)移(yí)動(dòng)的(dí)全地形車(chē)上(shàng)。相機在(zài)車輛(liàng)行(háng)進(jìn)過程中連(lián)續(xù)采集(jí)數據。

為適應(yīng)不(bù)同時段的光(guāng)照(zhào)變(biàn)化，係統會智能(néng)調整(zhěng)幀率(shuài)（40-50幀/秒(miǎo)），每個測(cè)量區塊平均(jūn)獲取710條掃描綫，總計(jì)約(yuē)63.9萬個光(guāng)譜(pǔ)像素點(diǎn)，。係統(tǒng)還配(pèi)備RTK校正功能(néng)，為所有(yǒu)采集(jí)數據提(tí)供厘米(mǐ)級(jí)的地理(lǐ)參考。

A Spanish team advanced the technology further by developing an innovative "on-the-go" hyperspectral imaging system. They mounted a visible-to-near-infrared hyperspectral camera (400-1000nm, 2.1nm spectral resolution) on an all-terrain vehicle moving at 5 km/h, enabling continuous data acquisition during operation.

To adapt to varying light conditions, the system automatically adjusts frame rates (40-50 fps). Each measurement block captures approximately 710 scan lines, totaling ~639,000 spectral pixels. The system also integrates RTK correction for centimeter-level georeferencing of collected data.

圖(tú)2 / Figure 2
：

使用安裝在(zài)全(quán)地(dì)形(xíng)車(ATV)上的(dí)攝像頭以 5 公(gōng)裏(lǐ)/小(xiǎo)時(shí)的速度(dù)進(jìn)行移(yí)動(dòng)高(gāo)光(guāng)譜(pǔ)成(chéng)像。通過推(tuī)掃式掃描(miáo)，從ATV的(dí)運動中獲(huò)得整個葡(pú)萄樹冠(guān)層(céng)的圖(tú)像，並用於(yú)估計葡萄成(chéng)分(fēn)。

On-the-go hyperspectral imaging with a camera mounted on an all-terrain vehicle (ATV) at 5 km/h. Images of the entire vine canopy were obtained from the ATV's motion, by push-broom scanning, and used for the estimation of grape composition.

數據(jù)處(chǔ)理上采(cǎi)用了(liǎo)支持向(xiàng)量機(SVM)算法(fǎ)，通過五(wǔ)折(zhē)交(jiāo)叉(chā)驗(yàn)證，對糖度的預(yù)測(cè)達到(dào)R²=0.91(RMSE=1.358°Brix)，對外(wài)部(bù)樣本的(dí)預(yù)測(cè)R²=0.92(RMSE=1.274°Brix)。對花(huā)青(qīng)素濃度的預測也取(qǔ)得(dé)了R²=0.72(交叉(chā)驗證)和R²=0.83(外(wài)部(bù)預測(cè))的良(liáng)好(hǎo)結果。實(shí)現了(liǎo)對葡(pú)萄的(dí)可(kě)溶性固形物(wù)和花青素(sù)濃(nóng)度(dù)的實(shí)時監(jiān)測(cè)。

計算(suàn)性(xìng)能方麵，在Intel Core i7處理器(qì)(16GB內存(cún))上(shàng)，處理(lǐ)36幅(fú)高光譜圖像平均需要5小時35分鐘，相(xiāng)當(dāng)於(yú)每列掃(sǎo)描(miáo)綫處理時間約0.79秒。而(ér)使用訓練好(hǎo)的SVM模(mó)型預測(cè)單個(gè)樣本(běn)僅需0.05秒，展現了(liǎo)良好的實用性能(néng)。

For data processing, the team employed support vector machine (SVM) algorithms. Through five-fold cross-validation, sugar content predictions reached R²=0.91 (RMSE=1.358°Brix), with external validation achieving R²=0.92 (RMSE=1.274°Brix). Anthocyanin concentration predictions also showed strong results: R²=0.72 (cross-validation) and R²=0.83 (external validation), enabling real-time monitoring of soluble solids and anthocyanins.

In terms of computational performance, processing 36 hyperspectral images on an Intel Core i7 processor (16GB RAM) averaged 5 hours and 35 minutes (~0.79 seconds per scan line). Trained SVM models required only 0.05 seconds per sample prediction, demonstrating practical usability.

圖(tú)3 / Figure 3
：

(a)基於RGB通道的高光譜圖(tú)像（為便於說明，對直(zhí)方(fāng)圖進行(háng)了(liǎo)歸(guī)一化）。 (b)像素光(guāng)譜(pǔ)與葡萄標(biāo)準光譜的R²值相(xiāng)關矩陣（應用σ=1.0高斯(sī)平滑核(hé)函(hán)數）。 (c)基於R²≥0.75閾(yù)值的葡萄像素分割(gē)結果圖(tú)。

(a) Hyperspectral image from a block in red, green and blue (RGB) channels (histogram normalised for the sake of illustration). (b) Correlation matrix with R² values between the pixel spectrum and a grape reference spectrum. A Gaussian smoothing was applied with σ = 1.0. (c) Image with segmented grape pixels (pixels in (b) whose R² ≥ 0.75). All the images were stretched in the horizontal axis for aesthetic purposes.

圖(tú) 4/ Figure 4
：

可(kě)溶(róng)性固(gù)體物模(mó)型的交(jiāo)叉驗(yàn)證(zhèng)和(hé)預(yù)測結果.

Regression plot for (a) fivefold cross validation (R² = 0.91; RMSE = 1.358) and (b) prediction results (R² = 0.92; RMSE = 1.274) for the TSS models.

圖5 / Figure 5
：

漿(jiāng)果(guǒ)花(huā)青素濃度(dù)的交叉(chā)驗證和(hé)預測模(mó)型的(dí)回歸(guī)圖(tú).

Regression plot for (a) fivefold cross validation (R² = 0.72; RMSE = 0.282) and (b) prediction results (R² = 0.83; RMSE = 0.211) for the anthocyanin concentration models.

『挑(tiāo)戰與未來展(zhǎn)望(wàng) / Challenges and Future Outlook 』

盡管仍需解(jiě)決環境(jìng)幹擾，數據處(chǔ)理速(sù)度等問題，這兩(liǎng)項(xiàng)研(yán)究證(zhèng)實(shí)了(liǎo)高光譜(pǔ)技(jì)術(shù)在田間(jiān)應用(yòng)的可行(háng)性(xìng)。

我們(mén)期待與更(gēng)多(duō)農(nóng)業機構(gòu)合(hé)作，提(tí)供專(zhuān)業(yè)的(dí)高(gāo)光(guāng)譜硬(yìng)件解決方(fāng)案(àn)。我(wǒ)們的(dí)設備能夠為客戶(hù)的(dí)研發(fā)團隊(duì)提(tí)供高質量(liáng)的原(yuán)始(shǐ)數(shù)據(jù)，兼容多(duō)種數據格(gé)式。隨(suí)著(zhuó)技術優化，這項(xiàng)技(jì)術(shù)有望成(chéng)為智(zhì)慧(huì)葡(pú)萄園的標準(zhǔn)配(pèi)置，為(wéi)葡(pú)萄(táo)和(hé)葡(pú)萄酒(jiǔ)產(chǎn)業帶來精(jīng)細化管理(lǐ)的新時(shí)代(dài)。

While challenges like environmental interference and processing speed remain, these studies confirm hyperspectral technology's field applicability.

We look forward to collaborating with agricultural institutions to provide professional hyperspectral hardware solutions. Our equipment delivers high-quality raw data in multiple compatible formats for research teams. As the technology evolves, it is poised to become a standard tool for smart vineyards, ushering in a new era of precision management for grape and wine production.

案例來(lái)源(yuán) / Source
：

1. Benelli, A., Cevoli, C., Ragni, L., & Fabbri, A. (2022). Reprint of: In-field and non-destructive monitoring of grapes maturity by hyperspectral imaging. Biosystems Engineering, 223(Part B), 200-208.

2. Gutiérrez, S., Tardaguila, J., Fernández-Novales, J., & Diago, M. P. (2018). On-the-go hyperspectral imaging for the in-field estimation of grape berry soluble solids and anthocyanin concentration. Australian Journal of Grape and Wine Research, 24(2), 127-133.