請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/90503
標題: | 應用機器學習建立台灣本土土壤保水曲線轉換方程式:以中部農業土壤為例 Establish machine learning-based localized pedotransfer functions of soil water retention curves:a case study of agricultural soils in central Taiwan. |
作者: | 馮博煜 Bo-Yu Fung |
指導教授: | 許少瑜 Shao-Yiu Hsu |
關鍵字: | 土壤轉換方程式,土壤保水曲線,排列重要性,不確定性分析,無母數自助法, pedotransfer function,soil water retention curve,permutation importance,uncertainty analysis,nonparametric bootstrap, |
出版年 : | 2023 |
學位: | 碩士 |
摘要: | 本研究以台灣中部農業土壤為範例,以土壤質地(砂、坋、黏含量)、總體密度與有機質含量作為解釋變數(或稱輸入特徵),應用多變量線性迴歸、隨機森林、類神經網路三種不同架構建立台灣本土土壤保水曲線的轉換方程式(pedotransfer function,PTF),並進一步分析PTF的輸入特徵重要性與不確定性。為了釐清不同輸入特徵對於PTF最終預測結果的影響力,本研究透過排列重要性的分析,得知PTF在基質勢能(matric potential)大於-0.1bar時,以總體密度作為最重要的輸入特徵;隨著基質勢能的減少總體密度重要性下降,坋粒與有機質含量重要性則逐漸上升。而為了評估非線性PTF的預測結果(不為常態分佈且具有較高偏度)之不確定性,本研究使用無母數自助法評估PTF的誤差95%信賴區間與95%預測區間。藉由衡量預測區間之實際涵蓋率,確認土壤含水量的實際涵蓋率皆介於95±1%,再次檢核無母數自助法可以有效評估非線性模型的不確定性。預測區間的建立除了提供PTF預測值的可靠度資訊外,此預測區間未來可用於檢驗土壤量測資料有效性的依據。最後,本研究成果也顯示三種不同架構的本土PTF,針對台灣土壤進行轉換的結果皆優於美國農業部開發之PTF─Rosetta3,再次確認建立台灣本土PTF的必要性。 This study takes agricultural soils in central Taiwan as an example and uses soil texture (sand, silt, clay content), bulk density, and organic matter content as explanatory variables (or input features) to develop pedotransfer function (PTF) for the soil water retention curve using three different frameworks: multiple linear regression, random forest, and artificial neural networks. The study further analyzes the uncertainty and feature importance of PTF. Through permutation importance analysis, the study reveals that bulk density is the most important feature when the matric potential is larger than -0.1 bar in PTF predictions. As the matric potential decreases, the importance of bulk density decreases while the importance of silt and organic matter content gradually increases. To evaluate the uncertainty of the nonlinear PTF predictions, which do not follow a normal distribution and have higher skewness, the study uses the nonparametric bootstrap method to assess the 95% confidence intervals of PTF’s error and 95% prediction intervals of PTF. By measuring the coverage probability of the prediction intervals, the study confirms that the coverage probability of soil water content is approximately 95±1%, validating the effectiveness of the nonparametric bootstrap method in assessing the uncertainty of nonlinear models. In addition to providing reliability information for PTF predictions, the establishment of prediction intervals can also be used as a basis for testing the validity of soil measurement data. Finally, the results of this study demonstrate that the three different frameworks of local PTF outperform the PTF developed by United States Department of Agriculture(USDA), Rosetta3, in converting Taiwanese soils, reaffirming the necessity of developing local PTF for Taiwan. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/90503 |
DOI: | 10.6342/NTU202301530 |
全文授權: | 同意授權(全球公開) |
顯示於系所單位: | 生物環境系統工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-111-2.pdf 此日期後於網路公開 2025-07-12 | 7.98 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。