應用高光譜影像檢測番茄水潛勢之研究

Abstract

高光譜影像技術屬於非破壞性檢測技術，已被廣泛使用在各產業領域，農業上之應用包括精準農業、作物栽培管理、農產品品質檢測等。番茄在每個生長期對水分的需求是不同的，供水過多或不足都會影響番茄植株的生長和產量。因此，在種植過程中必須進行精確的灌溉，以提高作物產量與品質。傳統上，土壤水分含量或葉片水潛勢被用來觀察植物水分狀況的指標。然而，這些方法準確性有限且耗時長，而且土壤水分屬於間接量測，現行葉片水潛勢屬於破壞性量測，難以在番茄生產中實施。因此有必要開發非破壞性且直接對植株本體的感測技術，本研究自行開發的移動式台車線上型高光譜影像檢測系統，使用了銦鎵砷（InGaAs）材質的高光譜影像相機，檢測波段範圍為900 - 1700 nm，並利用LabVIEW與MATLAB兩套程式進行系統軟硬體的整合。系統應用了波長、平場、空間等校正技術進行校準，確保系統穩定運作。利用線性判別分析，自動快速提取葉片圖像，識別準確率達到94.68%。標準常態變量（SNV）散射校正的數學處理用於消除散射失焦圖像引起的光譜變異。系統放置於台車上，在番茄作物的實際栽培場域，以線上水平方向移動並擷取光譜及影像的資訊，進行葉片本體水潛勢的檢測。水潛勢的量測實驗，採用露點微伏水潛勢測定儀（HR-33T），並使用了10個樣本室進行分時檢測，以NaCl標準檢測液建立0 ~ -2.971 MPa 範圍的水潛勢標準檢量線，作為樣本檢測計算葉片水潛勢的標準依據。兩批番茄葉片各169個及199個樣本進行水潛勢量測，其結果介於 -0.446 ~ -1.911 MPa，標準差為0.284 MPa，均落在標準檢量線可檢測的範圍內。葉片水潛勢的預測模型採用修正部分最小平方迴歸法（MPLSR）來建立，其分析結果分別為：169個樣本校正組標準誤差SEC = 0.129、校正組決定係數rc2 = 0.779、交叉驗證組標準誤差SECV = 0.171、交叉驗證組決定係數rcv2 = 0.610；199個樣本校正組標準誤差SEC = 0.201、校正組決定係數RSQ = 0.814、交叉驗證標準誤差SECV = 0.230、交叉驗證組決定係數1-VR = 0.755，這兩批樣本模型證實了非常好的預測能力結果。本研究亦經由所建立的MPLSR葉片水潛勢預測模型，計算出所拍攝的葉片影像各像素點上的水潛勢值，以假彩色方式呈現葉片的水潛勢分布圖，透過清晰易懂的可視化方式提供葉片水潛勢分布的直觀判別。本研究建立快速、簡易、非破壞性的葉片水潛勢量測系統及技術，有助於將智慧農業技術具體應用於農業生產技術之提升。

Hyper-spectral imaging technology is a non-destructive detection technology, and it has been widely used in various industrial fields. Agricultural applications include precision agriculture, crop cultivation management and quality evaluation of agricultural products. Tomatoes have different water requirements in each growing period. Excessive water use or insufficient water supply will affect the growth and yield of tomato plants. Therefore, precise irrigation control is required in the cultivation process to improve crop yield and quality. Traditionally, the soil moisture content or leaf water potential has been used as an indicator of plant water status. These methods, however, have limited accuracy and are time-consuming, soil moisture is an indirect measurement, and the current leaf water potential is a destructive measurement, which is difficult to implement in tomato production. Therefore, it is necessary to develop non-destructive and living body plant sensing technology. This study developed a mobile-carrier online type hyper-spectral imaging system, in which a hyper-spectral imaging camera made of indium gallium arsenide (InGaAs) with a detection wavelength range of 900 - 1700 nm was used, and programs using LabVIEW and MATLAB were adopted to integrate the systems. System pre-calibrations including wavelength correction, flat field correction and spatial correction are applied to ensure the stable operation of the system. Linear Discriminant Analysis was utilized to automatically and quickly extract the leaf images, with the recognition accuracy of 94.68% was achieved. The mathematical processing of Standard Normal Variate scattering correction was used to remove the spectral variations caused by the defocused leave images. The system can be easily installed on a carrier and moved to the actual cultivation field of the tomato crops. The spectra and image information are captured by horizontally online shooting of images of tomato crops, and the water potentials of the living body of leaves can be directly measured. Regarding the water potential experiments, DewPoint Microvoltmeter HR-33T was used and 10 sample chambers were also used for time-sharing purpose. NaCl standard test solution was used to establish a water potential standard calibration equations in the range of 0 ~ -2.971 MPa as the reference basis for calculating leaf water potentials. The water potential was measured on 169 and 199 samples of two batches of tomato leaves, and the results ranged from -0.446 to -1.911 MPa with a standard deviation of 0.284 MPa, both falling within the detectable range of the standard calibration line. The MPLSR (Modified Partial Least Square Regression) prediction model for tomato water potential was established, and the two regression results were : 169 samples calibration group SEC (Standard Error of Calibration) = 0.129, rc2 (coefficient of determination of calibration) = 0.779, cross-validation group SECV (Standard Error of cross-validation) = 0.171, rcv2 (coefficient of determination of cross-validation) = 0.610 ; 199 samples calibration group SEC (Standard Error of Calibration) = 0.201 , RSQ (coefficient of determination of calibration, R-square) = 0.814, cross-validation group SECV(Standard Error of cross-validation) = 0.230, 1-VR (coefficient of determination of cross-validation, one minus the variance ratio) = 0.755, these two batch sample models confirmed very good predictive power results. The water potential values at each pixel of the leaf images can be calculated through MPLSR model. The distribution of the water potential over the entire leaf area could be presented in a pseudo-color graph, in which intuitive judgment of leaf water potential distribution through clear and easy-to-understand visualization can be provided. In this research, a fast, easy and non-destructive leaf water potential measurement system and techniques have been developed, contribute to the specific application of smart agricultural technology to the improvement of agricultural production technology.