劉昊靈,張仲雄,陳 昂,浦育歌,趙 娟,2,3,胡 瑾,2,3
融合光譜形態(tài)特征的蘋果霉心病檢測(cè)方法
劉昊靈1,張仲雄1,陳 昂1,浦育歌1,趙 娟1,2,3,胡 瑾1,2,3※
(1. 西北農(nóng)林科技大學(xué)機(jī)械與電子工程學(xué)院,楊凌 712100;2. 農(nóng)業(yè)農(nóng)村部農(nóng)業(yè)物聯(lián)網(wǎng)重點(diǎn)實(shí)驗(yàn)室,楊凌 712100; 3. 陜西省農(nóng)業(yè)信息感知與智能服務(wù)重點(diǎn)實(shí)驗(yàn)室,楊凌 7121 00)
針對(duì)輕微霉心病和健康蘋果光譜差異較小,致使基于可見/近紅外特征光譜的檢測(cè)方法對(duì)輕微霉心病檢測(cè)準(zhǔn)確率較低的問題。該研究將光譜形態(tài)特征與光譜特征融合的方法引入霉心病模型構(gòu)建,建立了融合光譜形態(tài)特征的判別模型。以215個(gè)蘋果可見/近紅外光譜為樣本,分析了不同預(yù)處理和特征提取組合對(duì)建模效果的影響,并完成了光譜特征的提?。环治鼋】倒兔剐牟√O果平均光譜的差異性,提取波峰、波谷等差異明顯的光譜形態(tài)特征點(diǎn),對(duì)比波段比、波段差和歸一化強(qiáng)度差三類形態(tài)特征獲取方法;最終建立光譜形態(tài)特征參數(shù)和光譜特征融合的蘋果霉心病模型。試驗(yàn)結(jié)果表明,歸一化預(yù)處理后提取的特征光譜和歸一化強(qiáng)度差形態(tài)特征融合后模型判別準(zhǔn)確率最高,在支持向量機(jī)模型中訓(xùn)練集、測(cè)試集判別準(zhǔn)確率分別為98.6%和96.3%。特別是當(dāng)發(fā)病程度小于10%時(shí),該研究的判別模型準(zhǔn)確率高于95%,表明通過融合光譜形態(tài)特征可以提升輕微病變霉心蘋果的判別準(zhǔn)確率。
光譜;病害;蘋果霉心??;光譜形態(tài)特征;歸一化強(qiáng)度差;支持向量機(jī)
霉心病是蘋果的一種真菌病害,由于沒有外部癥狀出現(xiàn),這種病害在蘋果切開前無法被識(shí)別[1]。切開病果可見心室發(fā)霉或褐變腐爛,果心充滿粉紅色、灰綠色、黑褐色或白色霉?fàn)钗铮舨【黄菩氖冶跀U(kuò)展到心室外,則會(huì)引起果肉腐爛[2]。霉心病發(fā)病機(jī)理是由多種弱寄生菌組成的浸染過程較為復(fù)雜的復(fù)合型內(nèi)部病害,其中鏈格孢菌、鐮刀菌、單端孢等真菌會(huì)產(chǎn)生70多種具有不同化學(xué)結(jié)構(gòu)的次級(jí)代謝產(chǎn)物,其中部分成分有影響生育、引發(fā)癌癥以及減弱人體免疫等負(fù)面作用[3-5]。
近年來,可見/近紅外光譜、電子鼻[6-8]和振動(dòng)聲學(xué)[9]等多種方法被用于蘋果霉心病檢測(cè)。但電子鼻檢測(cè)方法需要等待空間內(nèi)氣體累積,其檢測(cè)時(shí)間一般在60~160 s,速度較慢;振動(dòng)聲學(xué)檢測(cè)方法僅對(duì)單一種類真菌造成的心室腐爛霉心病有效并且檢測(cè)準(zhǔn)確率不高,而霉心病是多種真菌造成的復(fù)合病變,其癥狀種類較多;核磁共振等其他檢測(cè)方式成本過高不能投入實(shí)際應(yīng)用。因此可見/近紅外透射光譜檢測(cè)法由于檢測(cè)速度快、準(zhǔn)確率高、可解釋性強(qiáng)、成本低的優(yōu)點(diǎn),已成為蘋果霉心病無損檢測(cè)中最熱門的方法。基于全光譜進(jìn)行蘋果霉心病判別證明了采用可見/近紅外透射光譜檢測(cè)霉心病的可行性[10],但這種方式光譜信息過于冗雜,同時(shí)設(shè)備成本較高,不利于實(shí)際應(yīng)用。雖然相繼出現(xiàn)使用小波變換[11]、連續(xù)投影算法(successive projections algorithm, SPA)[12]和競(jìng)爭(zhēng)自適應(yīng)重加權(quán)采樣(competitive adaptive reweighted sampling, CARS)[13]等方法進(jìn)行霉心病特征波長(zhǎng)提取的研究,其一定程度上減少特征波長(zhǎng)的數(shù)量,減少冗余光譜信息對(duì)建模效果的影響,但限制光譜特征數(shù)量的能力較弱,不能有效降低成本。為進(jìn)一步減少光譜特征數(shù)量并降低開發(fā)成本,也出現(xiàn)使用CARS和SPA組合的方法完成霉心病判別模型中特征波長(zhǎng)提取的研究[14-15]。但是由于蘋果果徑大小、果心位置等都會(huì)影響光譜傳播特性,從而影響模型判別效果,因此出現(xiàn)了考慮檢測(cè)方向差異的全局方向補(bǔ)償模型和考慮果徑大小對(duì)透射光譜影響的果徑修正模型[16-17]。其較經(jīng)典光譜分析方法一定程度上解決了果形差異導(dǎo)致霉心病檢測(cè)精度降低的問題。雖然該類方法和經(jīng)典光譜模型均對(duì)嚴(yán)重霉心病蘋果判別準(zhǔn)確率高,但對(duì)于輕微霉心病判別準(zhǔn)確率較低[18]。蘋果作為一種儲(chǔ)藏水果,能否在儲(chǔ)存前實(shí)現(xiàn)對(duì)輕微霉心病果的檢測(cè),就成為防治病害傳播、減少儲(chǔ)存損失的關(guān)鍵。因此迫切需要研究提高輕微霉心病蘋果的判別準(zhǔn)確率的方法。
光譜形態(tài)特征能夠充分利用光譜有效特征信息[19],目前利用光譜形態(tài)特征提高早期缺陷和成熟度檢測(cè)等品質(zhì)分級(jí)精度的方法[20-22]可以有效提升霉心病的判別準(zhǔn)確率。雖然光譜形態(tài)特征可以量化特征點(diǎn)之間的關(guān)系,但是單獨(dú)采用這種方法對(duì)特征光譜進(jìn)行處理會(huì)導(dǎo)致光譜原始信息的丟失,因此對(duì)判別準(zhǔn)確率的提升較為有限。
針對(duì)上述問題,本研究基于可見/近紅外光譜判別蘋果霉心病原理,提出一種將特征光譜和光譜形態(tài)特征融合的建模方法。通過判別精度選擇最佳特征光譜和光譜形態(tài)特征參數(shù),將特征光譜與光譜形態(tài)特征參數(shù)共同作為輸入訓(xùn)練模型,以提高可見/近紅外透射光譜判別蘋果霉心病的準(zhǔn)確率。
本文以紅富士蘋果為研究對(duì)象,于2020年11月在陜西省寶雞市扶風(fēng)縣發(fā)病率較高果園挑選發(fā)育良好、無外部損傷的紅富士蘋果215個(gè)。將蘋果運(yùn)回實(shí)驗(yàn)室后清洗干凈,使用標(biāo)簽紙逐一編號(hào)后在室溫(20 ℃,相對(duì)濕度為30%~40%)放置24 h[23]。
試驗(yàn)使用自主搭建的透射光譜采集平臺(tái),其結(jié)構(gòu)如圖1所示。試驗(yàn)儀器采用海洋光學(xué)的地物光譜儀(OFS-1100,Ocean Optics,USA),其采集光譜范圍350~1 100 nm,光譜分辨率1.3 nm;檢測(cè)光源為鋁制反射珠鹵素?zé)簦≦R111,Osram,USA);試驗(yàn)裝置安裝在一個(gè)密封的光屏蔽暗箱內(nèi),以防外部光干擾。試驗(yàn)過程光譜儀設(shè)置積分時(shí)間為100 ms;平均次數(shù)為5次;平滑度為10。在軟件中進(jìn)行去除暗噪聲、添加非線性矯正和雜散光矯正。另外蘋果內(nèi)部發(fā)霉部位并不規(guī)則,可能存在發(fā)霉部位偏離中心的情況,蘋果表面缺陷和其他病害也會(huì)對(duì)檢測(cè)造成影響,蘋果豎放時(shí)果萼和果柄也可能影響透射的光譜信息[24],因此為避免上述原因?qū)е鹿庾V缺失部分光譜信息,規(guī)定對(duì)每個(gè)樣本沿赤道方向每旋轉(zhuǎn)120°采集一次光譜,共獲得三組光譜信息[25]。使用Kennard Stone (K-S)算法將訓(xùn)練集與測(cè)試集按3∶1劃分[26]。
圖1 蘋果可見/近紅外光譜采集平臺(tái)
將蘋果樣本縱向切開,進(jìn)行蘋果霉心病果或健康果的分類。為了分析蘋果的發(fā)病程度,使用相機(jī)拍攝蘋果圖像獲取具體的蘋果病害信息,信息獲取過程如圖2所示。為保證圖像獲取環(huán)境相同,圖像拍攝時(shí)使用手機(jī)支架將手機(jī)相機(jī)固定在距黑色桌布12 cm的位置,使用光照計(jì)(LI-250A)將環(huán)境光強(qiáng)保持在235~243 lux,相機(jī)使用手機(jī)(小米11)后置1.08億像素?cái)z像頭。圖像處理在Photoshop軟件中進(jìn)行,根據(jù)圖像中蘋果剖面面積1和霉心面積2計(jì)算發(fā)病面積在剖面面積中的占比并將其定義為發(fā)病程度,以此評(píng)價(jià)霉心病蘋果的患病程度[27]。
圖2 蘋果病害信息獲取
不同的光譜預(yù)處理組合會(huì)對(duì)光譜特征提取和模型建模效果產(chǎn)生不同影響。S-G卷積平滑(savitzky-golay convolution smoothing, S-G)通過對(duì)移動(dòng)窗口內(nèi)的數(shù)據(jù)進(jìn)行多項(xiàng)式分解與最小二乘擬合可以有效去除光譜噪聲以提高光譜的信噪比;多元散射矯正(multiplicative scatter correction, MSC)和標(biāo)準(zhǔn)正態(tài)變換(standard normal variate transform, SNV)則可以解決光譜譜圖中參雜的固體顆粒密度、樣本折射率相關(guān)的問題;歸一化算法(normalization, NOR)可以有效避免某些重要特征權(quán)重較小的問題,在光譜預(yù)處理中也可以消除微小光程差異帶來的影響。因此,本文先對(duì)光譜進(jìn)行歸一化,然后分別選擇不再做處理、S-G平滑、MSC和SNV處理,建立4類預(yù)處理樣本集。
為提高判別模型的準(zhǔn)確率和運(yùn)行效率,對(duì)上述樣本集數(shù)據(jù)采用CARS與SPA結(jié)合的特征提取方法進(jìn)行降維處理。這種方法不僅能消除CARS提取的特征中連續(xù)和重疊部分,還能提高SPA的運(yùn)行速度,最大化壓縮數(shù)據(jù)維度,用更少的光譜信息表達(dá)更多的數(shù)據(jù)特征。本文首先使用CARS設(shè)置交叉驗(yàn)證次數(shù)10次,聚類次數(shù)25次對(duì)4類預(yù)處理結(jié)果進(jìn)行特征波長(zhǎng)提取,然后使用SPA限制最大特征數(shù)為20進(jìn)行二次提取,最后采用偏最小二乘判別分析(partial least squares discriminant analysis, PLS-DA)建立模型,以準(zhǔn)確率、召回率、特異性和特征數(shù)為指標(biāo)對(duì)特征提取結(jié)果進(jìn)行評(píng)價(jià),從而選擇建模效果最優(yōu)的預(yù)處理方法和特征光譜。
現(xiàn)有蘋果品質(zhì)檢測(cè)一般將其視為均勻介質(zhì),通過Lambert-Beer定律來描述光透過蘋果時(shí)的現(xiàn)象[28]。該定律認(rèn)為,在蘋果外觀物理因素一定的條件下,吸光度與吸收介質(zhì)的濃度和吸光系數(shù)有關(guān),所以霉心病蘋果的透射光強(qiáng)普遍低于健康蘋果。但試驗(yàn)表明,輕微的霉心部位在可見/近紅外光譜范圍內(nèi)透射光強(qiáng)度的衰減較小且不均勻,并且蘋果品質(zhì)、缺陷等其他因素會(huì)覆蓋輕微霉心病對(duì)透射光的影響甚至使光譜強(qiáng)度高于健康蘋果,導(dǎo)致僅采用特征光譜進(jìn)行輕微霉心病判別精度不高。因此,本研究從平均光譜中選擇5個(gè)波峰和波谷,使用波段比、波段差和歸一化強(qiáng)度差等波段運(yùn)算方式[29]獲取光譜形態(tài)特征(spectral shape features, SSF),建立PLS-DA判別模型對(duì)比準(zhǔn)確率,獲取最佳光譜形態(tài)學(xué)特征。
波段比(band ratio, BR)是2個(gè)波段光譜之間的比值[30],其不但可以有效地降低蘋果測(cè)量時(shí)造成的誤差,還可以增強(qiáng)波段之間的強(qiáng)度差異,提供一些任何單一波段無法得到的獨(dú)特信息[31]。BR數(shù)學(xué)表達(dá)式如式(1)所示,式中R和R分別是第和波段的光譜強(qiáng)度,BR是第和波段的光譜強(qiáng)度比值。
BR=R/R(1)
波段差(band difference, BD)則是2個(gè)波段光譜強(qiáng)度之間的差值,用于反應(yīng)光譜吸收峰的深度。BD數(shù)學(xué)表達(dá)式如式(2)所示,式中BD是第和波段的光譜強(qiáng)度差值。
BD=R?R(2)
歸一化強(qiáng)度差(normalized spectral intensity difference, NSID)是一種標(biāo)準(zhǔn)化的光譜指數(shù),該方法被應(yīng)用于西瓜的成熟度檢測(cè)[32],說明其具有一定發(fā)掘更深處光譜信息的作用。NSID數(shù)學(xué)表達(dá)式如式(3)所示,式中NSID是第和波段的光譜歸一化強(qiáng)度差,在遙感[33]和高光譜[34]中也被稱為歸一化植被指數(shù)。
支持向量機(jī)(support vector machine, SVM)算法可有效解決小樣本、非線性及高維模式識(shí)別的問題,因此本文采用SVM建立判別模型。采用特征光譜和BR、BD和NSID 3種光譜形態(tài)特征參數(shù)融合作為模型輸入,如式(4)所示,式中表示模型輸入,表示特征光譜,表示光譜形態(tài)特征參數(shù);以樣本標(biāo)簽(健康樣本為0,霉心病樣本為1)作為模型輸出。
={(,b)} (4)
為解決該類非線性分類問題,需通過引入核函數(shù)(,x)將低維線性不可分問題轉(zhuǎn)化為高維線性可分。最終得到的SVM判別模型式(5),式中表示樣本總數(shù)量,為L(zhǎng)agrange系數(shù),表示支持向量機(jī)的類別標(biāo)簽,為閾值。
本研究使用MATLAB中的fitcsvm二分類函數(shù)構(gòu)建SVM模型,并使用徑向基核函數(shù)(RBF)對(duì)樣本進(jìn)行高維度映射。為了確定最佳的框約束參數(shù)和內(nèi)核比例參數(shù),使用該函數(shù)的超參數(shù)優(yōu)化功能進(jìn)行自動(dòng)尋優(yōu)。
此外,為了驗(yàn)證SVM方法的優(yōu)越性,利用PLS-DA采用相同樣本集,使用25折交叉驗(yàn)證計(jì)算最佳潛在變量后建立模型進(jìn)行對(duì)比分析。
為驗(yàn)證建模效果,本文采用準(zhǔn)確率、特異性和召回率作為評(píng)價(jià)指標(biāo)來進(jìn)行模型好壞以及穩(wěn)健性的評(píng)估[35]。準(zhǔn)確率是指正確分類樣本占總樣本的比例,準(zhǔn)確率越高說明分類模型越好。在本研究中將霉心病樣本設(shè)為正例,召回率表示所有正例中被正確分類的比例,衡量了分類器對(duì)正例的識(shí)別能力,召回率越高說明對(duì)模型對(duì)霉心病果的檢出率更高。特異性表示所有負(fù)例中被正確分類的比例,衡量了分類器對(duì)負(fù)例的識(shí)別能力,特異性越高說明健康蘋果的誤判率越低。
本文在剔除蘋果光譜信息中異常樣本的基礎(chǔ)上,對(duì)550~900 nm波段內(nèi)光譜信息進(jìn)行分析。經(jīng)過統(tǒng)計(jì)后樣本總數(shù)215個(gè),其中健康蘋果111個(gè),霉心病樣本104個(gè)。得到霉心病樣本和健康樣本的光譜范圍如圖3表示,2個(gè)區(qū)域上下限分別代表霉心病蘋果光譜和健康蘋果光譜的各波段平均光譜±標(biāo)準(zhǔn)差。
圖3 健康和霉心病蘋果的光譜范圍
從圖3中可以看出霉心病蘋果光譜和健康蘋果光譜在各波段均有重合,尤其在650~700 nm的波谷中光譜幾乎完全重合,因此僅靠光強(qiáng)度對(duì)于這類霉心病樣本無法進(jìn)行有效判別。
為了避免光譜某些波段光強(qiáng)度過高或過低導(dǎo)致的特征權(quán)重不均衡的問題,必須先歸一化,然后再進(jìn)行S-G平滑、SNV、MSC等預(yù)處理,處理結(jié)果如圖4所示。
圖中分別是進(jìn)行NOR、NOR+SG、NOR+SNV和NOR+MSC預(yù)處理后特征提取的結(jié)果。其中NOR和NOR+SG均得到8維特征光譜,并且特征分布比較相似; NOR+SNV和NOR+MSC兩種方法分別得到11維和5維特征光譜。進(jìn)一步從圖中可以看出NOR和NOR+SG得到的特征光譜更符合光譜曲線的特征;NOR+SNV雖然得到特征數(shù)量較多,但在640、760、810 nm附近特征較為相似,并且相比另外3種方法缺少了710 nm附近波峰信息;NOR+MSC方法可以進(jìn)一步壓縮特征數(shù)量,但是從光譜曲線上看,丟失了640和810 nm附近波峰信息。為進(jìn)一步分析上述預(yù)處理方法的優(yōu)劣,使用PLS-DA建立判別模型對(duì)4種預(yù)處理方法進(jìn)行比較。利用不同處理方法得到的特征光譜建立PLS-DA判別模型,其結(jié)果如表1所示。
圖4 不同預(yù)處理方法下光譜特征提取結(jié)果
表1 不同預(yù)處理方式的PLS-DA模型訓(xùn)練結(jié)果
其中雖然NOR+SNV獲得特征數(shù)量最多,但是判別準(zhǔn)確率并不是最高,這可能是由于其缺少710 nm附近波峰信息,該波段是影響霉心病判別的主要特征波段[36];NOR+MSC獲得特征數(shù)量最少,并且判別準(zhǔn)確率最低;NOR和NOR+SG的特征較為相似,但是NOR預(yù)處理方式的判別準(zhǔn)確率更高。綜上所述,僅用NOR可以獲得更高的效益,因此采用NOR+CARS+SPA獲得的特征波長(zhǎng)建模。
平均光譜中波峰和波谷提取結(jié)果如圖5所示,分別為639、674、705、751和806 nm,以此5個(gè)波段作為提取光譜形態(tài)特征的特征波段。其中705 nm附近峰是影響霉心病的主要波段;806 nm附近峰是C-H鍵和O-H鍵的伸縮振動(dòng)引起的,該波段也被認(rèn)為和水份相關(guān)[37],而水果真菌感染與內(nèi)果壁鈣濃度有關(guān),鈣濃度變化是由果實(shí)水分導(dǎo)致的[38],因此800 nm吸收峰能夠反映霉心病的發(fā)病情況;639 nm附近峰值代表葉綠素B的吸收[39],674 nm附近波谷代表葉綠素A的降解[40],由于葉綠素與霉心病相關(guān)性較低,所以其相關(guān)波段在波段運(yùn)算中用于校正690 nm后波段的光學(xué)性質(zhì)[41];751 nm附近的波谷值一般被認(rèn)為和806 nm同樣作為水分的吸收峰,但該波段與果核的光學(xué)性質(zhì)有關(guān)[42],因此在果核發(fā)病的蘋果霉心病中也應(yīng)考慮該波段的影響。
利用式(1)~(3)對(duì)上述5個(gè)波段進(jìn)行運(yùn)算獲得4個(gè)BR變量、4個(gè)BD變量和4個(gè)NSID變量。將BD、BR、NSID輸入建立的PLS-DA霉心病判別模型,其結(jié)果如表2所示。
圖5 平均光譜的波峰和波谷
對(duì)比結(jié)果可以發(fā)現(xiàn),NSID的判別準(zhǔn)確率最高,BD、BR的判別準(zhǔn)確率較低。雖然3種波段運(yùn)算方法的準(zhǔn)確率均在70%以上,但建模效果仍略低于特征光譜直接建模。說明光譜形態(tài)特征雖然可以反映出霉心病相關(guān)的特征,但是也丟失了一部分特征信息。為了進(jìn)一步提高判別準(zhǔn)確率,需要將4種光譜形態(tài)特征添加到特征光譜中建模分析,評(píng)價(jià)其建模效果。
表2 不同波段運(yùn)算方法的PLS-DA模型準(zhǔn)確率
根據(jù)上述分析,為了對(duì)比引入不同形態(tài)特征對(duì)建模效果的影響,本文對(duì)比了在特征光譜中不做處理、融合BR、融合BD和融合NSID后,使用SVM和PLS-DA方法分別建立判別模型的準(zhǔn)確率,如表3所示。
表3 不同建模方法下建模效果對(duì)比
從表3中可知,融合光譜形態(tài)特征對(duì)蘋果霉心病的判別能力有很大提升,并且在SVM中效果要優(yōu)于PLS-DA。在三種光譜形態(tài)特征融合中,融合NSID的判別效果最好,訓(xùn)練集準(zhǔn)確率98.1%,召回率和特異性分別為97.0%和98.5%,測(cè)試集中準(zhǔn)確率96.3%,召回率和特異性分別為100%和94.9%,其次是融合BR,最后是融合BD。其中,融合BR和融合BD的效果與表2中結(jié)果相反,說明BR特征相比BD特征更能與特征光譜互補(bǔ)。另外融合NSID的模型召回率明顯高于融合其他兩種光譜形態(tài)特征的模型,說明NSID特征對(duì)判別準(zhǔn)確率的提升主要體現(xiàn)在對(duì)霉心病蘋果的識(shí)別能力。因此,本文選擇NSID和特征光譜融合建立SVM霉心病判別模型。
為了驗(yàn)證使用本文方法構(gòu)建模型的性能,利用相同的樣本集,將融合光譜形態(tài)特征模型與使用霉心病相關(guān)特征波段建模所得的特征波段模型、檢測(cè)方向補(bǔ)償模型[16]和果徑修正模型[17]進(jìn)行對(duì)比,最終結(jié)果如表4所示。表中可以看出果徑修正模型、檢測(cè)方向補(bǔ)償模型和融合形態(tài)特征模型均取得了高于特征波段建模的結(jié)果。其中融合形態(tài)特征模型的判別準(zhǔn)確率最高,表明融合光譜形態(tài)特征的方法對(duì)霉心病的判別效果更好。
表4 蘋果霉心病判別模型對(duì)比
為進(jìn)一步分析本文方法對(duì)不同發(fā)病程度霉心病判別準(zhǔn)確率的提升效果,對(duì)上述4種方法的霉心病誤判樣本和發(fā)病程度進(jìn)行統(tǒng)計(jì)如圖6所示。在霉心病中最輕微的癥狀是僅在果核內(nèi)部發(fā)病并且沒有侵染果肉,而根據(jù)統(tǒng)計(jì)規(guī)律果核面積一般占蘋果橫截面積的6%,因此將發(fā)病程度小于6%作為輕微霉心病的標(biāo)準(zhǔn);另外在一些相關(guān)文獻(xiàn)中將發(fā)病程度小于10%定義為輕微霉心病[27],因此定義發(fā)病程度在6%~10%作為為略微嚴(yán)重的輕微霉心??;另外發(fā)病程度10%~14%為一般霉心病,而發(fā)病程度大于14%發(fā)病面積時(shí)蘋果已經(jīng)發(fā)生嚴(yán)重霉變,這種情況最容易判別??梢园l(fā)現(xiàn)隨著發(fā)病程度的減少,霉心病的判別準(zhǔn)確率呈下降趨勢(shì)。當(dāng)發(fā)病程度大于14%時(shí),除特征波段建模以外,其他三種模型都能夠?qū)崿F(xiàn)霉心病精準(zhǔn)判別;當(dāng)發(fā)病程度大于10%且小于14%時(shí),只有融合光譜形態(tài)特征模型能夠100%判別,果徑修正模型和檢測(cè)方向補(bǔ)償模型雖然出現(xiàn)誤判,但準(zhǔn)確率仍高于特征波段建模;當(dāng)發(fā)病程度小于10%但大于6%時(shí),特征波段建模的判別準(zhǔn)確率的下降非常嚴(yán)重,另外兩種模型的判別準(zhǔn)確率也出現(xiàn)下滑,只有融合光譜形態(tài)特征模型的判別準(zhǔn)確率明顯高于其他幾種模型且仍能達(dá)到95.7%;當(dāng)發(fā)病程度小于6%時(shí),特征波段建模僅有33.3%的判別準(zhǔn)確率,檢測(cè)方向補(bǔ)償模型判別準(zhǔn)確率為75.0%,果徑修正模型判別準(zhǔn)確率為83.3%,而融合光譜形態(tài)特征模型判別準(zhǔn)確率仍能達(dá)到95.8%。本文模型的準(zhǔn)確率相對(duì)特征波段建模和果徑修正模型分別提高了62.5和12.5個(gè)百分點(diǎn)。結(jié)果表明,當(dāng)發(fā)病程度小于10%時(shí),融合光譜形態(tài)特征的的模型對(duì)于判別準(zhǔn)確率的提升非常顯著,說明融合光譜形態(tài)特征的模型可以有效表達(dá)出霉心病蘋果和健康蘋果間光譜的差異性,從而提高模型識(shí)別輕微霉心病蘋果的能力。
圖6 不同蘋果霉心病判別模型對(duì)不同發(fā)病程度判別的準(zhǔn)確率分布
為了進(jìn)一步分析本方法局限性,對(duì)本方法錯(cuò)判的91、95號(hào)樣本進(jìn)行分析,其光譜如圖7所示。91與95號(hào)樣本發(fā)病程度分別為4.9%和6.9%,觀察剖面圖像發(fā)現(xiàn),其屬于果核內(nèi)部出現(xiàn)少量菌絲的輕微霉心病癥狀[43]。不僅如此,樣本維管束附近果肉均有呈透明玻璃狀趨勢(shì),這是典型的蘋果水心病特征[44-45]。由于在水心病組織中光的透射率高于無癥狀組織[46],因此在樣本中輕微霉心病導(dǎo)致的透射光強(qiáng)衰減低于水心導(dǎo)致的透射光強(qiáng)增加的情況下,各波段光譜強(qiáng)度甚至高于健康樣本平均光譜。上述結(jié)果表明水心病組織對(duì)光譜的影響會(huì)覆蓋霉心病組織影響,導(dǎo)致無法從光譜中獲得霉心病組織的光譜形態(tài)特征,從而出現(xiàn)誤判現(xiàn)象。在上述兩種病害并存的情況下,引入新的修正因子提升模型判別精度將成為下一步研究的重點(diǎn)。
圖7 融合光譜形態(tài)特征模型中判錯(cuò)樣本分析
另一方面,由于不同品種蘋果間的果實(shí)形狀、顏色、內(nèi)部品質(zhì)差異較大,導(dǎo)致不同品種之間蘋果的光學(xué)性質(zhì)差異明顯,尤其光譜特征峰的漂移嚴(yán)重,因此現(xiàn)有模型并不能在品種中通用。因此面對(duì)不同品種蘋果時(shí)重新獲取樣本按本研究步驟建立具有針對(duì)性的模型,或考慮使用遷移學(xué)習(xí)的方法將本研究模型本研究模型推廣至其他品種的蘋果。
此外本研究中健康樣本和霉心病樣本數(shù)量分布均勻,若健康樣本多于霉心病樣本,會(huì)導(dǎo)致建模結(jié)果特異性增加而召回率降低,若霉心病樣本多于健康樣本,會(huì)導(dǎo)致召回率增加而特異性降低,不平衡的樣本分布會(huì)導(dǎo)致對(duì)某一種樣本判別能力降低本研究模型的準(zhǔn)確率。因此后續(xù)研究中考慮采用生成對(duì)抗網(wǎng)絡(luò)(generative adversarial network, GAN)或人工少數(shù)類過采樣(synthetic minority over-sampling technology, SMOTE)[26]等生成樣本的方式保證樣本數(shù)量平衡,并與本研究方法結(jié)合可能會(huì)進(jìn)一步提高判別準(zhǔn)確率。
為了提高輕微霉心病的檢測(cè)精度,本文提出一種融合光譜形態(tài)特征的蘋果霉心病判別模型建模方法。在選擇光譜歸一化為預(yù)處理方法,使用競(jìng)爭(zhēng)自適應(yīng)重加權(quán)采樣和連續(xù)投影算法結(jié)合完成特征光譜提取的基礎(chǔ)上,分別對(duì)光譜形態(tài)特征波段做波段比、波段差、歸一化強(qiáng)度差的波段運(yùn)算后與特征光譜融合作為輸入建立判別模型進(jìn)行對(duì)比。最終發(fā)現(xiàn)使用歸一化強(qiáng)度差和特征光譜融合建立的支持向量機(jī)模型效果最佳,其訓(xùn)練集準(zhǔn)確率達(dá)到98.1%、測(cè)試集準(zhǔn)確率達(dá)到96.3%。將該模型與特征波段建模、果徑修正模型和檢測(cè)方向補(bǔ)償模型進(jìn)行對(duì)比發(fā)現(xiàn),融合光譜形態(tài)特征的模型判別效果更好。尤其當(dāng)發(fā)病程度小于6%時(shí),其相對(duì)特征波段建模準(zhǔn)確率提高了62.5個(gè)百分點(diǎn),相對(duì)以往準(zhǔn)確率最高的果徑修正模型提高了12.5個(gè)百分點(diǎn),證明本文方法能有效提升輕微霉心病蘋果的判別精度。
[1] ALI S, ABBASI P A, REHMAN S, et al. First report of moldy core of minneiska apples from new zealand caused by alternaria arborescens[J]. Plant Disease, 2021, 105(9): 2719-2719.
[2] 周兆永,何東健,張海輝,等. 基于深度信念網(wǎng)絡(luò)的蘋果霉心病病害程度無損檢測(cè)[J]. 食品科學(xué),2017,38(14):297-303.
ZHOU Zhaoyong, HE Dongjian, ZHANG Haihui, et al. Non-destructive detection of moldy core in apple fruit based on deep belief network[J]. Food Science, 2017, 38(14): 297-303. (in Chinese with English abstract)
[3] LIU Y, YAMDEU J H G, GONG Y Y, et al. A review of postharvest approaches to reduce fungal and mycotoxin contamination of foods[J]. Comprehensive Reviews in Food Science and Food Safety, 2020, 19(4): 1521-1560.
[4] LOPEZ P, VENEMA D, DE R T, et al. Occurrence of alternaria toxins in food products in the netherlands[J]. Food Control, 2016, 60: 196-204.
[5] NTASIOU P, MYRESIOTIS C, KONSTANTINOU S, et al. Identification, characterization and mycotoxigenic ability of Alternaria spp. causing core rot of apple fruit in Greece[J]. International Journal of Food Microbiology, 2015, 197: 22-29.
[6] GUO Z M, GUO C, CHEN Q S, et al. Classification for penicillium expansum spoilage and defect in apples by electronic nose combined with chemometrics[J]. Sensors, 2020, 20(7): 2130-2144.
[7] JIA W S, LIANG G, TIAN H, et al. Electronic nose-based technique for rapid detection and recognition of moldy apples[J]. Sensors, 2019, 19(7): 1526-1537.
[8] 楊晨昱,袁鴻飛,馬惠玲,等. 基于傅里葉近紅外光譜和電子鼻技術(shù)的蘋果霉心病無損檢測(cè)[J]. 食品與發(fā)酵工業(yè),2021,47(7):211-216.
YANG Chenyu, YUAN Hongfei, MA Huiling, et al. Non-destructive detection of apple moldy core based on FT-NIR and electronic nose technology[J]. Food and Fermentation Industries, 2021, 47(7): 211-216. (in Chinese with English abstract)
[9] ZHAO K, ZHA Z H, Li H, et al. Early detection of moldy apple core based on time-frequency images of vibro-acoustic signals[J]. Postharvest Biology and Technology, 2021, 179: 111589.
[10] SHENDEREY C, SHMULEVICH I, ALCHANATIS V, et al. Nirs detection of moldy core in apples[J]. Food and Bioprocess Technology, 2010, 3(1): 79-86.
[11] ZHOU Z Y, LEI Y, SU D, et al. Detection of moldy core in apples and its symptom types using transmittance spectroscopy[J]. International Journal of Agricultural and Biological Engineering, 2016, 9(6): 148-155.
[12] 張本華,錢長(zhǎng)錢,焦晉康,等. 基于介電特性與SPA-SVR算法的長(zhǎng)稻含長(zhǎng)率檢測(cè)方法[J]. 農(nóng)業(yè)工程學(xué)報(bào),2019,35(18):237-244.
ZHANG Benhua, QIAN Changqian, JIAO Jinkang, et al. Rice moisture content detection method based on dielectric properties and SPA-SVR algorithm[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2019, 35(18): 237-244. (in Chinese with English abstract)
[13] 唐海濤,孟祥添,蘇循新,等. 基于CARS算法的不同類型土壤有機(jī)質(zhì)高光譜預(yù)測(cè)[J]. 農(nóng)業(yè)?程學(xué)報(bào),2021,37(2): 105-113.
TANG Haitao, MENG Xiangtian, SU Xunxin, et al. Hyperspectral prediction on soil organic matter of different types using CARS algorithm[J]. Transactions of the C hinese Society of Agricultural Engineering (Transactions of the CSAE), 2021, 37(2): 105-113. (in Chinese with English abstract)
[14] WANG Z L, FAN S X, WU J Z, et al. Application of long-wave near infrared hyperspectral imaging for determination of moisture content of single maize seed[J]. Spectrochimica Acta Part a-Molecular and Biomolecular Spectroscopy, 2021, 254: 119666.
[15] KAMRUZZAMAN M, KALITA D, AHMED T M, et al. Effect of variable selection algorithms on model performance for predicting moisture content in biological materials using spectral data[J]. Analytica Chimica Acta, 2022, 1202: 339390.
[16] ZHANG K X, JIANG H Z, ZHANG H C, et al. Online detection and classification of moldy core apples by VIS-NIR transmittance spectroscopy[J]. Agriculture-Basel, 2022, 12(4): 489-506.
[17] TIAN S J, ZHANG J H, ZHANG Z X, et al. Effective modification through transmission Vis/NIR spectra affected by fruit size to improve the prediction of moldy apple core[J]. Infrared Physics & Technology, 2019, 100: 117-124.
[18] 張海輝,田世杰,馬敏娟,等. 考慮直徑影響的蘋果霉心病透射光譜修正及檢測(cè)[J]. 農(nóng)業(yè)機(jī)械學(xué)報(bào),2019,50(1):313-320.
ZHANG Haihui, TIAN Shijie, MA Minjuan,et al. Detection method of moldy core in apples using modified transmission spectrum based on size of fruit[J]. Transactions of the Chinese Society for Agricultural Machinery, 2019, 50(1): 313-320. (in Chinese with English abstract)
[19] 黃敏,朱曉,朱啟兵,等.基于高光譜圖像的玉米種子特征提取與識(shí)別[J].光子學(xué)報(bào),2012,41(7):868-873.
HUANG Min, ZHU Xiao, ZHU Qibing, et al. Morphological characteristics of maize seed extraction and identification based on the hyperspectral image[J]. ACTA Photonica Sinica, 2012, 41(7): 868-873(in Chinese with English abstract)
[20] TIAN X, ZHANG C, Li J B, et al. Detection of early decay on citrus using LW-NIR hyperspectral reflectance imaging coupled with two-band ratio and improved watershed segmentation algorithm[J]. Food Chemistry, 2021, 360: 130077.
[21] LEE H, KIM M S, JEONG D, et al. Detection of cracks on tomatoes using a hyperspectral near-infrared reflectance imaging system[J]. Sensors, 2014, 14(10): 18837-18850.
[22] ZHANG M S, SHEN M S, PU Y G, et al. Rapid identification of apple maturity based on multispectral sensor combined with spectral shape features[J]. Horticulturae, 2022, 8(5): 361-376.
[23] ZHANG M S, SHEN M S, LI H, et al. Modification of the effect of maturity variation on nondestructive detection of apple quality based on the compensation model[J]. Spectrochimica Acta Part a-Molecular and Biomolecular Spectroscopy, 2022, 267: 120598.
[24] TIAN S J, ZHANG M S, LI B, et al. Measurement orientation compensation and comparison of transmission spectroscopy for online detection of moldy apple core[J]. Infrared Physics & Technology, 2020, 111: 103510.
[25] XIA Y, FAN S X, LI J B, et al. Optimization and comparison of models for prediction of soluble solids content in apple by online Vis/NIR transmission coupled with diameter correction method[J]. Chemometrics and Intelligent Laboratory Systems, 2020, 201: 104017.
[26] ZHAO F Y, DU G R, HUANG Y. Exploring the use of near-infrared spectroscopy as a tool to predict quality attributes in prickly pear (Rosa roxburghii Tratt) with chemometrics variable strategy[J]. Journal of Food Composition and Analysis, 2022, 105: 104225.
[27] ZHANG Z X, LIU H L, CHEN D Y, et al. SMOTE-based method for balanced spectral nondestructive detection of moldy apple core[J]. Food Control, 2022, 141: 109110.
[28] ZUDE M, PFLANZ M, SPINELLI L, et al. Non-destructive analysis of anthocyanins in cherries by means of Lambert-Beer and multivariate regression based on spectroscopy and scatter correction using time-resolved analysis[J]. Journal of Food Engineering, 2011, 103(1): 68-75.
[29] LI L, PENG Y K, LI Y Y, et al. Rapid and low-cost detection of moldy apple core based on an optical sensor system[J]. Postharvest Biology and Technology, 2020, 168: 111276.
[30] 江洪,何國(guó)金,黃海明,等. 基于波段比模型的地形調(diào)節(jié)植被指數(shù)組合算法構(gòu)建與驗(yàn)證[J]. 農(nóng)業(yè)工程學(xué)報(bào),2017,33(5):156-161.
JIANG Hong, HE Guojin, HUANG Haiming, et al. Construction and validation of combination model of topography-adjusted vegetation index based on band-ratio model[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2017, 33(5): 156-161. (in Chinese with English abstract)
[31] 蔡健榮,王建黑,陳全勝,等. 波段比算法結(jié)合高光譜圖像技術(shù)檢測(cè)柑橘果銹[J]. 農(nóng)業(yè)工程學(xué)報(bào),2009,25(1):127-131.
CAI Jianrong, WANG Jianhei, CHEN Quansheng, et al. Detection of rust in citrus by hyperspectral imaging technology and band ratio algorithm[J] Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2009, 25(1): 127-131. (in Chinese with English abstract)
[32] JIE D F, ZHOU W H, WEI X. Nondestructive detection of maturity of watermelon by spectral characteristic using NIR diffuse transmittance technique[J]. Scientia Horticulture, 2019, 257: 108718.
[33] 袁媛,瑚波,武興厚,等.基于植被指數(shù)的夏玉米不同生育期葉綠素含量遙感估算[J].中國(guó)農(nóng)學(xué)通報(bào),2015,31(15):254-259.
YUAN Yuan, HU Bo, WU Xinghou, et al. Remote Sensing estimation of chlorophyll content in different growth periods of summer corn based on vegetation index[J]. Chinese Agricultural Science Bulletin, 2015, 31(15): 254-259. (in Chinese with English abstract)
[34] 吾木提·艾山江,尼加提·卡斯木,陳晨,等.基于多維高光譜植被指數(shù)的冬小麥葉面積指數(shù)估算[J].農(nóng)業(yè)機(jī)械學(xué)報(bào), 2022,53(5):181-190.
UMUT H, NIJAT K, CHEN Chen, et al. Estimation of winter wheat LAI based on multi-dimensional hyperspectral vegetation indices[J]. Transactions of the Chinese Society for Agricultural Machinery, 2022, 53(5): 181-190. (in Chinese with English abstract)
[35] POURDARBANI R, SABZI S, KALANTARI D, et al. Automatic non-destructive video estimation of maturation levels in Fuji apple (Malus Malus pumila) fruit in orchard based on colour (Vis) and spectral (NIR) data[J]. Biosystems Engineering, 2020, 195: 136-151.
[36] 張海輝,陳克濤,蘇東,等. 基于特征光譜的蘋果霉心病無損檢測(cè)設(shè)備設(shè)計(jì)[J]. 農(nóng)業(yè)工程學(xué)報(bào),2016,32(18):255-262.
ZHANG Haihui, CHEN Ketao, SU Dong, et al. Design of nondestructive detection device for moldy core in apples based on characteristic spectrum[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2016, 32(18): 255-262. (in Chinese with English abstract)
[37] WANG G, LIU Y, LI X, et al. Differences in optical properties of fruit tissues between stone fruits and citrus fruits[J]. Infrared Physics & Technology, 2020, 112: 103599.
[38] LEVIN A G, YERMIYAHU U, DORON I, et al. The role of calcium concentration in the endocarp wall of apple fruit in the development of core rot[J]. Crop Protection, 2019, 120: 67-74.
[39] ROBBE V B, BEN A, RODRIGO W, et al. Effect of maturation on the bulk optical properties of apple skin and cortex in the 500 to 1850 nm wavelength range[J]. Journal of Food Engineering, 2017, 214: 79-89.
[40] STEFAN A, KONNI B, STEFFEN N, et al. Determining the optical properties of apple tissue and their dependence on physiological and morphological characteristics during maturation. Part 1: Spatial frequency domain imaging[J]. Postharvest Biology and Technology, 2021, 181: 111647.
[41] ZUDE M. Comparison of indices and multivariate models to non-destructively predict the fruit chlorophyll by means of visible spectrometry in apple fruit[J]. Analytica Chimica Acta, 2003, 481(1): 119-126.
[42] SEIFERT B, ZUDE M, SPINELLI L, et al. Optical properties of developing pip and stone fruit reveal underlying structural changes[J]. Physiol Plantarum, 2015, 153: 327-336.
[43] 趙康,查志華,李賀,等. 基于聲振信號(hào)對(duì)稱極坐標(biāo)圖像的蘋果霉心病早期檢測(cè)[J]. 農(nóng)業(yè)工程學(xué)報(bào),2021,37(18):290-298.
ZHAO Kang, ZHA Zhihua, LI He, et al. Early detection of moldy apple core using symmetrized dot pattern images of vibro-acoustic signals[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2021, 37(18): 290-298. (in Chinese with English abstract)
[44] KASAI S, ARAKAWA O. Antioxidant levels in watercore tissue in ‘Fuji’ apples during storage[J]. Postharvest Biology and Technology, 2010, 55(2): 103-107.
[45] GAO Z F, JAYANTY S, BEAUDRY R, et al. Sorbitol transporter expression in apple sink tissues: Implications for fruit sugar accumulation and watercore development[J]. Journal of the American Society for Horticultural Science, 2005, 130(2): 261-268.
[46] CHANG H, WU Q, TIAN H, et al. Non-destructive identification of internal watercore in apples based on online VIS/NIR spectroscopy[J]. Transactions of the Asabe, 2020, 63(6): 1711-1721.
Detection method for apple moldy cores based on spectral shape features
LIU Haoling1, ZHANG Zhongxiong1, CHEN Ang1, PU Yuge1, ZHAO Juan1,2,3, HU Jin1,2,3※
(1., 712100,; 2.,,, 712100,;3., 712100,)
Moldy core is one of the most serious fungal diseases in apples. The visible/near infrared spectroscopy (VIS-NIR) technique has been a common-used approach to distinguish the apple moldy core. However, the better discriminant can be only confined to the severely diseased apples in the existing VIS-NIR, due to the smaller spectral difference between the mild mold core and healthy apples. It is a high demand to detect the mild mold core for early warning during apple production. In this study, an improved discriminant model was established to detect the apple moldy core using the spectral shape features, in order to significantly improve the detection accuracy. 215 well-developed red Fuji apples without external damage were selected from the orchard in Fufeng County, Baoji City, Shaanxi Province, China, in November 2020. The VIS-NIR (350-1100nm) information was first collected from these apples. The images were then captured from the cutting apples. The degree of moldy-core was determined to calculate the ratio of the mold core area to the apple profile before image pretreatment. The discriminative accuracy was firstly compared with the savitzky-golay convolution smoothing (S-G) after normalization (NOR), Multiplicative scatter correction (MSC) after NOR, and standard normal variate transform (SNV) after NOR. Secondly, the feature bands were extracted from the images using the combination of competitive adaptive reweighted sampling (CARS), and Successive projections (SPA). Thirdly, five peaks and valleys (at the wavelength of 639, 674, 705, 751, and 806 nm) were extracted from the average spectrum for the typical shape features. Band ratio (BR), band difference (BD), and normalized spectral intensity difference (NSID) were then analyzed to determine the spectral shape features (SSF) parameters with the highest discriminant accuracy. Finally, the optimal model was obtained between the partial least squares discriminant analysis (PLS-DA) and support vector machine (SVM). Results show that the NOR spectral pretreatment performed the best to extract the characteristic spectrum, whereas the NSID was the best among the three SSF parameters. The SVM model presented the highest discriminative accuracy with the training set of 98.6%, and the test set of 96.3%. Four models were used to evaluate the performance of model identification in the different degrees of moldy-core, including the build the modal with characteristic band, the correct spectrum with apple diameter, compensate with the direction of detection, and merge the spectral shape seatures. Once the degree of moldy-core was greater than 10%, the accuracies of these models were improved significantly, except for thebuild the modal with characteristic band. When the degree of moldy-core was less than 10%, only the Merged Spectral Shape Features Model performed a high discriminant accuracy of higher than 95%, which was 43.5 percentage points higher than the build the modal with characteristic band. In the case of the moldy-core degree less than 6%, the discrimination accuracy of merge the spectral shape features reached 95.8%, which was 62.5 percentage points, and 12.5 percentage points higher than the build the modal with characteristic band, and the correct spectrum with apple diameter, respectively. Consequently, the discrimination model merged with the NISD in the input of the apple mold core can be expected to greatly improve the discrimination accuracy, particularly for the mild mold core. The improved model merged with the spectral shape features can be an effective way to accurately discriminate the apple moldy core.
spectroscopy; disease; mild mold core; spectral shape features; normalized spectral intensity difference; support vector machine
10.11975/j.issn.1002-6819.202210038
O657.33; S436.611
A
1002-6819(2023)-01-0162-09
劉昊靈,張仲雄,陳昂,等. 融合光譜形態(tài)特征的蘋果霉心病檢測(cè)方法[J]. 農(nóng)業(yè)工程學(xué)報(bào),2023,39(1):162-170.doi:10.11975/j.issn.1002-6819.202210038 http://www.tcsae.org
LIU Haoling, ZHANG Zhongxiong, CHEN Ang, et al. Detection method for apple moldy cores based on spectral shape features[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2023, 39(1): 162-170. (in Chinese with English abstract) doi:10.11975/j.issn.1002-6819.202210038 http://www.tcsae.org
2022-10-08
2022-11-16
陜西省科技重大專項(xiàng)(2020ZDZX03-05-01);國(guó)家自然科學(xué)基金項(xiàng)目(31701664)
劉昊靈,研究方向?yàn)檗r(nóng)產(chǎn)品無損檢測(cè)與裝備研發(fā)。Email:2332835570@nwafu.edu.cn
胡瑾,博士,教授,博士生導(dǎo)師,研究方向?yàn)檗r(nóng)業(yè)信息感知與智能決策。Email:hujin007@ nwsuaf.edu.cn