請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/28978
標題: | 應用智慧型光譜資訊分析於蔬菜植株氮含量檢測之研究 Analysis of Nitrogen Content in Vegetables Using Intelligent Spectral Information |
作者: | Chia-Tseng Chen 陳加增 |
指導教授: | 陳世銘(Suming Chen) |
關鍵字: | 鳳京小白菜,甘藍苗,近紅外光,光譜影像,類神經網路,資訊熵,基因遺傳演算法, Chinese Mustard,Cabbage Seedlings,Near Infrared (NIR),Spectral Image,Artificial Neural Network (ANN),Information Entropy (IE),Genetic Algorithm (GA), |
出版年 : | 2007 |
學位: | 博士 |
摘要: | 近年來,應用遙測式光譜檢測技術於植株生理狀態分析已相繼被提出。本研究為能發展具備線上即時且非破壞性的光譜影像遙測系統,以作為蔬菜作物在生長過程中氮營養狀態之監測。採用近紅外光分光光度計(NIRS 6500, FOSS NIRSystems)和自行開發之高光譜影像系統,進行蔬菜植株之光譜資料量測;以典型光譜檢測模式,包括逐步型多重線性迴歸(step-wise multilinear regression, SMLR)與修正型部份最小平方迴歸(modified partial least square regression, MPLSR),初步探討植株氮含量之光譜檢測能力;再以具備機器學習能力之演算法,包括類神經網路模式(artifical neural network, ANN)、實數基因遺傳演算法(real genetic algorithm, RGA)及光譜資訊熵(information entropy, IE)分析,建立智慧型多光譜氮含量檢測模式,以增進蔬菜植株中氮含量之檢測能力。
本研究第一階段以三種不同氮肥栽培之113個小白菜「鳳京」葉片粉末樣本,進行乾粉氮含量之光譜檢測模式分析。光譜經由微分處理後,其光譜值與氮含量濃度間之相關係數,在1400-2450 nm波段範圍內,其相關係數絕對值可達0.9以上,故以此波段範圍為氮含量乾粉光譜分析之有效特徵波段。在氮含量光譜檢測模式分析中,以平滑化併一次微分之四波長(2124 nm、2240 nm、1666 nm和 1632 nm)SMLR模式預測能力為最佳(SEC = 2.059 mg/g, rc = 0.991, SEV = 2.131 mg/g, rv = 0.990),優於採用相同光譜前處理之最佳MPLSR模式,顯示僅使用少數波長建構之光譜檢測模式,在無水份光譜吸收的干擾條件下,仍可勝於MPLSR。因此,可利用此一植株葉片氮含量之光譜檢測模式,取代傳統耗時費工的濕式化學分析。此外,本研究成果指出,若進一步考慮將光譜檢測技術應用於現場植株生長之氮肥營養管理,可以採用短波近紅外光攝影機搭配液晶可調濾波器拍攝植株高光譜影像,以二次微分併平滑化之450-1000 nm光譜波段資訊,建構植株高光譜影像之MPLSR氮含量檢測模式,以供生產現場管理之用。 第二階段以五種不同氮營養處理之甘藍苗鮮葉,進行鮮葉反射吸收光譜之量測。為能滿足未來發展多光譜影像檢測系統以進行作物氮含量之遙測,本研究以特徵波長分析和光譜檢測模式進行探討。在特徵波長選擇中,分別針對全波段(400-2500 nm)、矽質CCD攝影機感光波段(450-950 nm),利用SMLR進行鮮葉氮含量之光譜特徵波長分析。在本研究中,特別提出之具備交互學習能力之類神經網路模式(ANN-CL),以發展具有更佳預測能力之氮含量光譜檢測模式。在考慮以矽質CCD攝影機和市售可選之帶通濾鏡規格,建構較具應用潛力之多光譜影像遙測系統,以SMLR分析可得四特徵波長(490, 570, 600, and 680 nm)組合之結果,經由ANN-CL進行光譜檢測模式分析後,可得之結果為rc=0.93、SEC=0.873 %和SEV=0.960 %,可有效提升原來SMLR之預測能力高達約15 %(原SEV=1.099 %)。此外,利用ANN-CL之四特徵波長所得之分析結果,已可比擬SMLR使用全波段七特徵波長時,所建立之光譜檢測模式分析結果(rc=0.94, SEC=0.985 %, SEV=0.993 %)。因此,利用ANN-CL使用四特徵波長(490, 570, 600, and 680 nm)組合,可以建構出具備實用性之多光譜影像遙測系統,以預測甘藍種苗之氮含量。 第三階段自行開發完成實驗室型高光譜影像拍攝系統,結合了兩組矽質CCD攝影機與LCTF分光器(VIS and VNIR),擷取波長範圍為410-1090 nm的甘藍苗植被高光譜影像資料,以進行甘藍苗植株樣本氮含量分析。在高光譜影像分析中,因具備良好的空間解析能力,得以準確利用簡單之二元化背景分離程序,找出植被前景與背景之影像資訊,輔以同時拍攝之校正色板所得之六組校正色塊影像資訊,在選擇適當校正色塊組合之平均灰度值與反射率資訊後,可建立灰度與反射率轉換函式,以正確估算植被之反射光譜資料。利用有效資訊熵指標分析植株氮含量之特徵波長後,其第一和第二特徵波長分別為650 nm和690 nm,與氮含量與光譜值相關性分析結果互異。而第三特徵波長為530 nm,而氮含量與光譜值相關性分析所得之特徵波長為520 nm,兩者差異不大。氮含量與光譜值相關性分析所得之第四特徵波長470 nm,在有效資訊熵指標分析後,仍小於波長760 nm和900 nm之估算結果。因此,在有效資訊熵指標之特徵波長分析結果中,可得五個特徵波長組合結果(分別為650, 690, 520, 760, 900 nm)。以原始植株高光譜影像資料之典型光譜檢測模式分析結果而言,可以在使用五個因子數進行MPLSR分析後,其標準校正誤差可收斂至6.20 mg/g,而其標準交互驗證誤差也可收斂至7.64 mg/g;就SMLR之模式預測能力與特徵波長選擇結果而言,可以利用三波長組合(470, 710, 1080 nm)進行線性迴歸模式分析,獲得最佳之SMLR光譜檢測模式,其SEC可收斂至7.55 mg/g,而SEV為8.13 mg/g。 進一步將氮含量與光譜值相關性分析(LC)、光譜資訊熵分析(IE)與SMLR所得之三組特徵波長組合,分別作為實數基因遺傳演算法(RGA)與交互學習類神經網路模式(ANN-CL)之智慧型光譜檢測模式輸入變數,以建立較佳之多光譜影像植株氮含量檢測模式。在RGA之分析中,因其藉由隨機產生基因組合與優良基因交配等機制,具備良好的能力以避免模式落入區域最小誤差,因此利用IE所得之五個特徵波長組合(520, 650, 690, 760, 900 nm),可得具備較佳之模式預測能力(SEV=7.79 mg/g)。利用ANN-CL進行智慧型光譜檢測模式分析,由校正樣本組中以比例3/4的樣本數,隨機選擇每一訓練回合之學習樣本進行訓練後,可以同時的提升校正樣本組與驗證樣本組之氮含量預測能力(SEC= 6.47 mg/g, SEV= 5.76 mg/g)。 綜上所論,本研究以近紅外光與光譜影像技術,結合智慧型模式之演算法,已成功建立良好蔬菜植株之氮含量光譜檢測模式,未來將可進一步地利用少數特徵波長建構多光譜影像遙測系統,實際應用於溫室之蔬菜作物氮營養監測,藉由提供準確之植株氮含量分析結果,發展適時適量之精準氮營養管理模式。 Using spectral remote sensing to monitor the physiological status during growth has been attempted in the recent studies. In the work, the near infrared spectrophotometer (NIRS 6500, FOSS NIRSystems Inc.) and the hyper-spectral imaging system developed in this study were used to measure and analyze the reflectance spectra of vegetables in order to provide the basis for the future development of the on-line non-destructive remote sensing system for monitoring the nitrogen content of vegetable crops. The typical calibration models, including step-wise multilinear regression (SMLR) and modified partial least square regression (MPLSR), were adopted to examine the prediction performance of plant nitrogen content by using the spectral data firstly. Furthermore, the machine learning algorithms, including artificial neural network (ANN), real genetic algorithm (RGA), and information entropy (IE), were adopted to develop the intelligence-based calibration models to improve the prediction accuracy of calibration models. In the first part of this dissertation, 113 samples of Chinese mustard (Brassica rapa L. var. chinensis (Rupr.) Olsson) were cultured by three different nitrogen fertilization treatments, and the reflectance spectra of leaves in terms of powder form were used to develop the calibration models. The results show that derivative treatments can reduce the noises of spectral shift caused by the particle sizes, and the significant wavelengths with high correlation coefficient ( |r| > 0.9 ) appear in the selected significant spectral band (1400-2450 nm). Regarding the nitrogen prediction accuracy, the SMLR model with smooth and first derivative pre-treatments and four significant wavelengths (2124, 2240, 1666, and 1632 nm) gives the best results (SEC = 2.059 mg/g, rc = 0.991, SEV = 2.131 mg/g, rv = 0.990). The results point out the SMLR model with a few wavelengths as inputs can be better than MPLSR model when spectral information is without water absorbance interference. Moreover, the SMLR model could be used to replace the time-consuming wet chemical method, such as Kjeldahl method, to analyze the nitrogen content in vegetable leaves. The results also indicate that a hyper-spectral imaging system, constructed of silicon CCD cameras and liquid crystal tunable filters (LCTF) using MPLSR method with the smooth and second derivative spectral information in range of 450 to1000 nm, could be used as the aids for nitrogen fertilization management of vegetable growth in the field. In the second part, fresh leaves of cabbage seedlings (Brassica oleracea L.) after fertilizations with 5 different concentrations are used to measure the reflectance absorbance spectra. To develop a multi-spectral imaging system for remote sensing of the nitrogen content of crops, the significant wavelengths and calibration models were carefully evaluated in this study. The significant wavelengths in full band (400-2500 nm) and a selected band (450-950 nm), which is suitable for silicon CCD cameras, were investigated. Significant wavelengths for estimating nitrogen content of cabbage seedling leaves were first determined by SMLR analysis. A proposed ANN model with cross-learning scheme (ANN-CL) was further developed to increase the prediction accuracy. To comply with the design of a practical multi-spectral imaging system using silicon CCD cameras and commercially available bandpass filters, an ANN-CL model with four inputs of spectral absorbance at 490, 570, 600, and 680 nm was developed. The calibration results (rc = 0.93, SEC = 0.873%, and SEV = 0.960%) reduce the SEV about 15% when compared with the SMLR method with four wavelengths (SEV = 1.099%). In addition, the results are comparable to that of SMLR with seven wavelengths (rc = 0.94, SEC = 0.806%, and SEV = 0.993%) in the full band. These results indicate that the ANN model with cross-learning using spectral information at 490, 570, 600, and 680 nm could be used to develop a practical remote sensing system to predict nitrogen content of cabbage seedlings. In the third part, the self-developed hyper-spectral imaging system, constructed from two sets of CCD cameras and liquid crystal tunable filters (LCTF, VIS and VNIR), were used to grab the spectral images of cabbage seedlings in the wavelength band of 410-1090 nm. In the analysis of hyper-spectral images, the region of seedling canopy was precisely extracted by image segmentation, which was dealt with a simply binary procedure, due to the fine spatial resolution of images. To calibrate and transfer the gray values of seedling canopy to the reflectance absorption, the six standard gray-blocks were used. The first and second significant wavelengths, analyzed by the information entropy (IE) index, are 650 nm and 690 nm, which are mutual different to the linear correlation (LC) analysis between nitrogen content and spectral data. The third significant wavelength of IE analysis is 530 nm, which is similar to 520 nm of LC. However, the fourth significant wavelength of LC is 470 nm, whose index value of IE is less than the wavelengths of 760 nm and 900 nm. The significant wavelengths of IE analysis are including 650, 690, 520, 760, and 900 nm. In the results of hyper-spectral calibration model analysis by using raw spectral data, MPLSR with six factors reduces the values of SEC and SECV to 6.20 mg/g and 7.64 mg/g respectively. Besides, the SMLR with three significant wavelengths (470, 1080, and 710 nm) gives the best results (SEC=7.55 mg/g, SEC=8.13 mg/g) by using simply linear equation. The different significant wavelengths sets of LC, IE and SMLR are used as input data of the intelligence-based calibration models of RGA and ANN-CL to improve the prediction accuracy of nitrogen content analysis. Regarding the RGA analysis, the genetic population was generated randomly and the best fitness genetic population was kept to generate the next generation by crossover and mutation, and the global minimum of error was achieved. Therefore, the RGA calibration model with five significant wavelengths set (650, 690, 520, 760, and 900 nm) of IE is obtained with the good prediction results (SEV=7.79 mg/g). Moreover, the intelligence-based calibration model of ANN-CL with 3/4 sample selection ratio of the calibration set, using the same significant wavelengths set of RGA model, reduces the SEC to 6.47 mg/g and SEV to 5.76 mg/g effectively. As a conclusion, the study has successfully developed nitrogen content prediction models using multi-spectra data of vegetable crops by integrating the near infrared, spectral images technology and artificial intelligence algorithms. With these research results, the remote sensing system with a multi-spectral imager could be developed for monitoring the nitrogen status of greenhouse crops in the future. The information of crops nitrogen status is useful for the precision management of nitrogen fertilization. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/28978 |
全文授權: | 有償授權 |
顯示於系所單位: | 生物機電工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-96-1.pdf 目前未授權公開取用 | 5.01 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。