Please use this identifier to cite or link to this item:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/50187
Title: | 以熱影像技術檢測番茄水分逆境之研究 Evaluation of Tomato Water Stress by Thermal Imaging Technology |
Authors: | Pin-Chih Fang 方品智 |
Advisor: | 陳世銘(Suming Chen) |
Keyword: | 熱影像,影像分割技術,作物水分逆境指數,牛番茄,葉片水勢, infrared thermography,semantic segmentation,crop water stress index,tomato,leaf water potential, |
Publication Year : | 2020 |
Degree: | 碩士 |
Abstract: | 農農產業為臺灣用水量最高的產業,而造成農業用水量過大的其中一個原因,便是無法方便地定義植株水分逆境,並針對該逆境情形進行澆灌。而如何快速、方便的檢測植株的水分逆境,便成為農產業提高生產效率的重要課題。前人研究,顯示藉由葉片溫度判斷植株的缺水情形為可行的,因此本研究使用熱影像模組以及熱影像儀開發檢測系統,並蒐集量測環境溫度、照度及葉片溫度,建立預測牛番茄葉片水勢之預測模型。 本研究使用熱像儀於溫室內量測定植後一到三個月的牛番茄植株,並建立預測牛番茄葉片水勢(Leaf Water Potential, LWP)之預測模型。預測模型參考作物水分逆境指數(Crop Water Stress Index, CWSI)的設計,該預測模式參考環境中澆灌良好植株之表面溫度(The Temperature of Wet Surface)及無蒸散作用植株之表面溫度(The Temperature of Dry Surface)。本研究中使用原版之CWSI模型預測葉片水勢之結果,r2=0.63,標準校正誤差(Standard Error of Calibration, SEC)=0.219 MPa;而若於預測模型加入環境照度及溫度資訊,得以改善於溫室應用作物水分逆境指數預測葉片水勢之預測能力,其結果r2值為0.776,SEC=0.16 MPa,故此模型具有很好的預測葉片水勢的能力,且熱影像儀具有在溫室內預測植株水分逆境的可能性。試驗中亦使用多重線性迴歸(Multivariate Linear Regression, MLR)及人工類神經網路(Artificial Neural Network, ANN)的模型預測葉片水勢,其中以ANN的預測結果最佳,其r2 = 0.88,SEC=0.113 MPa.。而熱影像模組檢測系統之預測結果遠不如熱影像儀檢測系統,其原因為熱影像模組之量測穩定性尚不足,其預測葉片水勢結果之SEC=0.316,幾乎為使用相同預測模型的熱影像儀檢測系統之兩倍。 熱影像處理之部分,本研究參考全卷積神經網路(Fully Convolutional Network, FCN)分割熱影像中的葉片位置,再藉由葉片溫度、環境溫度及照度預測葉片水勢。60張熱影像用於訓練葉片語意分割(Semantic Segmentation)模型,在預測能力上有很不錯的表現,在訓練樣本中像素級別(Pixel-wise)的預測準確度達到99.2%,在驗證樣本中準確度亦達到98.3%,本研究證實應用深度學習之演算法於熱影像處理的可行性。 使用葉溫及環境參數預測植株缺水情形為可行的,且在預測模型中加入照度資訊是不可或缺的。研究中顯示熱影像儀可以準確的預測葉片水勢;熱影像模組部分,雖然目前無法準確定量植株的缺水情形,但由於其較為平價,若未來檢測準確度及穩定性得以提升,應用於農產業界也是指日可待。研究中亦驗證使用全卷積神經網路分割熱影像中葉片位置資訊有很好的效果,使得熱影像檢測水分逆境在自動化檢測系統中更具可行性。 Agriculture is the industry with the highest water consumption in Taiwan. One of the contributing factors to the excessive water consumption in the agriculture industry is that the adoption of sensing technology is often limited for plant water deficit in scheduling precise irrigation strategy. Therefore, it is important to develop a remote sensing technology which can measure water deficit efficiently. Previous studies have shown the feasibility of evaluating crop water stress using leaf temperature. Hence, the purpose of this study is to develop a measuring system with thermography and thermal camera module, coupled with leaf temperature, ambient temperature and illuminance. The developed system is applied to predict the leaf water potential (LWP) of tomatoes of 1-3 months after planting. The design of the prediction model utilizes the crop water stress index (CWSI), with both temperatures of dry and wet surfaces are required for the calculation of CWSI and improving the accuracy of water status prediction. To be more specific, the temperature of the wet surface represents the leaf temperature of well-watered plants; while the temperature of the dry surface indicates the leaf temperature of non-evaporation plants. By using the developed thermography system, the obtained result of LWP prediction, referring to the CWSI, is shown to be linearly related to the crop water stress index with coefficient of determination (r^2) = 0.63 and standard error of calibration (SEC) = 0.219. Furthermore, when the factor of environment illuminance is added to the modified CWSI, the performance of the LWP prediction has improved to r^2=0.776 and SEC = 0.16. Additionally, the multiple linear regression, nonlinear regression and artificial neural network are also applied to develop the leaf water potential prediction model. Among all, the artificial neural network provides the best prediction result, with r^2=0.88 and SEC = 0.113 are obtained. On the other hand, due to the instability and sensitivity of the thermal camera module, the prediction of LWP using the thermal camera module gives a less satisfactory result, with SEC = 0.316 is obtained. The obtained SEC value is almost two times larger than the prediction result using the thermography system. In this study, the deep learning algorithm is also applied in the thermal image processing. The fully convolutional network is used to extract leaf positions from the thermal image before calculating the leaf temperature. The extracted leaf temperature and environment parameter are used to predict the LWP. A total of 60 thermal images has been used to develop the semantic segmentation model. The pixel-wise accuracies of 99.2% and 98.3% are obtained for training and validation datasets, respectively. The attained accuracy indicates the beneficial potential of applying the fully convolutional network in thermal imaging processing. The feasibility of integrating the ambient temperature, illuminance and leaf temperature to predict the LWP has been shown in this study. Despite the thermal camera module could not effectively predict the LWP based on the current stability, the competitive advantage of low price may popularize its use in the agriculture industry, if the stability of the thermal camera module can be improved in near future. Also, the fully convolutional network works well in leaf temperature extraction. The findings in this study have thus suggested the potential utility of thermal images in automatic detection of LWP in the agriculture industry. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/50187 |
DOI: | 10.6342/NTU202002955 |
Fulltext Rights: | 有償授權 |
Appears in Collections: | 生物機電工程學系 |
Files in This Item:
File | Size | Format | |
---|---|---|---|
U0001-1108202014574700.pdf Restricted Access | 6.13 MB | Adobe PDF |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.