應用溫室內多功能監測系統於甘藍種苗生長性狀判別之研究

Abstract

本研究延續使用溫室內多功能監測系統，對不同生長階段的甘藍種苗擷取R、G、B及NIR波段影像，同時記錄苗齡、鮮物重和乾物重，以建立多光譜影像檢量模式。以室內種苗樣本多光譜影像之檢量模式來說，使用影像平均灰階、植被指數及紋理特徵作為輸入參數，建立了簡單多重線性迴歸、逐步多重線性迴歸、逐步迴歸之類神經網路及主成分分析之類神經網路共四種模式，其中以逐步迴歸之類神經網路具有最佳預測能力，其Rc、Rp、SEC、SEP及RPD在預測種苗苗齡為0.98、0.96、1.11、1.92及3.61；在預測種苗鮮物重為0.98、0.98、0.32、0.36及4.36，在預測種苗乾物重為0.99、0.96、0.02、0.05及3.61。 在溫室中擷取多光譜影像時，為避免現場光度變化造成影像曝光過度或不足，研究中建立了攝影機分割畫面之自動曝光演算法，並以LabVIEW 7.1版軟體撰寫控制程式，藉由設定影像範圍及灰階區間，以曝光控制演算法求得理想電子快門及增益值，進而擷取良好品質之影像。曝光控制性能測試方面，設定任意灰階上下限為10的範圍，該演算法皆可得到理想電子快門及增益值參數；時間響應方面，電子快門在任何灰階設定範圍狀況下，皆可在3秒鐘內完成控制，增益值方面，控制時間隨著灰階設定範圍與原影像平均灰階之差距成正比趨勢。 以溫室內多光譜影像而言，藉由空間校正、灰階校正及影像縫合等影像處理技巧得到苗床平面縫合影像，可瞭解目前苗床上苗盤分佈位置，並以NDVI空間分佈，予以四種不同灌溉水量，如此溫室內多功能監測系統在不同的時間，藉由擷取苗床平面影像及環境因子分析，可建立苗床位置之灌溉水量表，提供溫室內噴灑系統作為灌溉依據，以達到精準栽培之最終目標。

This study keep on the application of multi-function monitoring system for greenhouse production. To investigate different growth stage, fresh matter weight and dry matter weight of cabbage seedling, this research estalishs analytic model, by using R,G, B and NIR image. This study uses multiple linear regression, stepwise multiple linear regression, stepwise artificial neural network and principle component analysis artificial neural network with the input of image gray level average, vegetation indice and texture from cabbage seedlings samples. The best prediction model is stepwise artificial neural network. For prediction of seedings growth day, Rc=0.98, Rp=0.96, SEC=1.11, SEP=1.92 and RPD=3.61. For prediction of seedings fresh matter weight, Rc=0.98、Rp=0.98、SEC=0.32、SEP=0.36 and RPD=4.36. For prediction of dry matter weight, Rc=0.99、Rp=0.96、SEC=0.02、SEP=0.05 and RPD=3.61. To avoid the situation of image over-exposure or under-exposure, this study uses LabVIEW 7.1 software to establish an automatic exposure algorithm for camera shutter and gain control when grapping multi-image in greehouses. By set up image region of interest and gray level range the algorithm always find the optimal shutter and gain to grap image of good quality.In performance test, the algorithm works in any gray level range. In time response, it spends 3 seconds to finish shutter control. The execution time for gain control is proportional to the difference between initial image gray level and the setting. The whole image from spatial calibration, gray-level calibration and image stitchment can provide tray location information.With the NDVI spatial variation, there are four kinds of irrigation policies corresponding four NDVI ranges. For the purpose of precision agriculture in greenhouses, the monitoring system set up different amount of irrigation water for different cabbage seeglings and transmit this information to sprayer system