Please use this identifier to cite or link to this item:
標題: 毛豆合格莢產量之氣象估計模式的研究
Toward a Weather Model for Predicting the Marketable Pod Yield of Vegetable Soybean
作者: 林隆新
Lin, Lung-Hsin
關鍵字: 毛豆
vegetable soybean
marketable pod
predictor variables
ridge regression
principal components regression
partial least squares regression
relative prediction error
出版社: 農藝學系
摘要: 本研究之目的在開發一個可供估計台灣南部地區毛豆之合格莢產量的氣象模式。在1998至2000年之間,於屏東市的高雄區農業改良場內,以毛豆品種高雄五號進行一系列播種期×收穫期的田間試驗,總共獲得74筆毛豆合格莢產量的數據。田間試驗期間同時測記了日均溫、日照時數、日射量及降雨量等氣象變數。由於氣象數據過於龐大,我們先將毛豆的整個生育期間劃分為營養生長期與生殖生長期兩個階段,或劃分為營養生長期、開花著莢期及莢果充實期三個階段,再計算各生長階段內之日數、平均氣溫、平均日照時數、平均日射量及平均降雨量做為建立合莢格產量估計模式的基本推估變數。其次,我們將整套數據分裂為兩部分,較早獲得的42筆數據做為開發模式用的建模數據,其餘32筆數據則做為俟後測試模式之估計準確性的驗證數據。模式中的推估變數包含前述各生長階段內之日數、平均氣溫、平均日照時數、平均日射量及平均降雨量的一次項、二次項及彼此間的乘積項。由於這些推估變數彼此間的高度相關,我們嘗試以脊回歸法、主成分回歸法及淨最小平方和回歸法來解決這種共線的困擾。結果顯示,將整個生育期劃分為三個生長階段時,所求得的模式雖然對建模數據的配適程度相當高,但在以驗證數據測試時卻無法合理地估計合格莢產量。就三種求配的程序而言,脊回歸法的表現最差。當我們把整個生育期劃分為兩個生長階段時,主成分回歸法與淨最小平方和回歸法都能獲得相當好的結果,求配所得的模式不僅對建模數據有很高的配適度,在以驗證數據進行測試時也呈現相當準確的估計能力:估計的合格莢產量與其實測值之間相關高達0.85,而估計值的最大相對誤差則在22%以下。
Summary This study aims to develop a weather model for predicting the marketable pod yield of vegetable soybean in South Taiwan. A total of 74 observations on the marketable pod yield of Kaohsiung No. 5, a soybean variety bred specially for vegetable usage, were obtained in a series of sowing date × harvest date field experiments conducted at the Kaohsiung District Agricultural Improvement Station located in Pingtung City during the years of 1998 through 2000. Meanwhile, daily mean air temperature, sunshine hour, solar radiation and precipitation were measured during the time of field experiments. To manage the huge meteorological data, the whole growing period of each crop was divided into two stages (vegetative growth stage and reproductive growth stage), or three stages(vegetative growth stage, flowering-podding stage, and pod-filling stage). Thereby the duration of each growth stage and the mean of each of the four meteorological variables over each growth stage were taken as the candidate predictors for the marketable pod yield. The data was split into two portions, i.e., the 42 earlier observations were used as training data set to develop the models, and the rest 32 observations were reserved as test data set to check the prediction accuracy of the fitted models. Second-degree polynomial models in the mentioned predictors were fitted to the training data. The methods of ridge regression, principal components regression, and partial least squares regression were employed to cope with the collinearity among the predictor variables. The result shows that the models fitted with the whole growing period being divided into three stages were not able to predict the marketable pod yield in the test data, although their fitness to the training data were rather satisfactory. Among the three fitting procedures, ridge regression failed to give any model with reasonable prediction ability. With the whole growing period being divided into two stages, both principal components regression and partial least squares regression gave models with not only very good fitness to the training data but also faily high accuracy in predicting the marketable yield of the test data. The correlation coefficient between the predicted marketable pod yield and its observed value was as large as 0.85, and the maximum of the relative prediction error was as small as 22%.
Appears in Collections:農藝學系



Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.