2026-07-08
【學術亮點-頂級期刊論文】利用機器學習迴歸方法對多光譜無人機影像進行分析，估算乾旱異質性農田的精細尺度土壤含水量和小麥產量
字體大小
小
中
大
Intelligent Service: Large-scale Agricultural AI Models
【Department of Civil Engineering / Ming-Der Yang / Tenured Distinguished Professor】
智慧服務：可大規模擴展之農業AI 模型【土木工程學系楊明德終身特聘教授】
上架日期2026-07-07
【Department of Civil Engineering / Ming-Der Yang / Tenured Distinguished Professor】
智慧服務：可大規模擴展之農業AI 模型【土木工程學系楊明德終身特聘教授】
| 論文篇名 | 英文：Estimating fine-scale soil moisture content and wheat yield in drought-heterogeneous farmlands using machine learning regressions on multispectral UAV imagery 中文：利用機器學習迴歸方法對多光譜無人機影像進行分析，估算乾旱異質性農田的精細尺度土壤含水量和小麥產量 |
| 期刊名稱 | Computers and Electronics in Agriculture (指標清單期刊) |
| 發表年份, 卷數, 起迄頁數 | 2026,253, no.112145 |
| 作者 | Ming-Der Yang(楊明德), Tung-Ching Su(蘇東青)*, Yu-Chun Hsu(許鈺群) |
| DOI | 10.1016/j.compag.2026.112145 |
| 中文摘要 | 隨著精準農業的發展，土壤含水量（SMC）作為影響小麥產量的關鍵因素，需要在田間尺度上進行高效且近實時的監測。本研究利用無人機（UAV）多光譜影像擷取的土壤線指數，估算了台灣金門兩塊土壤含水量空間分佈差異顯著（即農田乾旱異質性）的農田的精細尺度土壤含水量和小麥產量。透過比較傳統線性迴歸與多種機器學習方法（包括支持向量迴歸（SVR）、梯度提升迴歸（GBR）和卷積神經網路迴歸（CNNR）），識別了不同小麥生長階段影響土壤含水量和產量的關鍵指數。結果表明，當農田乾旱異質性顯著時，土壤含水量和小麥產量均可得到較好的解釋。使用垂直植被指數進行線性迴歸，土壤含水量的最佳決定係數（p < 0.01）為0.65；使用歸一化植被指數進行線性迴歸，小麥產量的最佳決定係數為0.93。研究發現，農田中土壤含水量（SMC）較高的區域會對小麥生長產生負面影響。在機器學習模型中，基於選定指數（分別為VAPDI和NDVI）的GBR和CNNR模型對兩個試驗農田的小麥總產量預測準確率分別達到了98.0%和72.2%。跨不同生長週期的時間驗證表明，這些模型具有跨年小麥產量預測的潛力。與預期預測模式的偏差可視為產量異常的初步指標，早期產量預測展現出良好的應用前景。總而言之，這些發現為早期產量預測提供了初步證據，並為優化灌溉時間和加強人工智慧驅動的農業決策支援系統的開發提供了寶貴的見解。 |
| 英文摘要 | With the development of precision agriculture, soil moisture content (SMC), a key factor affecting wheat yield, requires efficient and near real-time monitoring at the field scale. This study applies soil line indices derived from unmanned aerial vehicle (UAV) multispectral imagery to estimate fine-scale SMC and wheat yield in two farmlands with distinct spatial SMC distributions (i.e., farmland drought heterogeneity) in Kinmen, Taiwan. By comparing traditional linear regression with several machine learning approaches, including support vector regression (SVR), gradient boosting regression (GBR), and convolutional neural network regression (CNNR), key indices contributing to SMC and yield estimation across different wheat growth stages were identified. The results indicate that both SMC and wheat yield can be reasonably explained when significant farmland drought heterogeneity is present. Linear regression achieved optimal coefficients of determination (p < 0.01) of 0.65 for SMC using the perpendicular vegetation index and 0.93 for wheat yield using the normalized difference vegetation index. Areas with higher SMC within a farmland were found to negatively affect wheat growth. Among the machine learning models, GBR and CNNR based on the selected indices (VAPDI and NDVI, respectively) achieved estimation accuracies of 98.0 % and 72.2 %, respectively, for total wheat yield across the two experimental farmlands. Temporal validation across different growing cycles suggests the potential for cross-year wheat yield prediction. Deviations from expected estimation patterns may be considered preliminary indicators of abnormal yield conditions, and early-stage yield estimation shows promising potential. Overall, these findings provide preliminary evidence for early-stage yield prediction and offer valuable insights for optimizing irrigation timing and enhancing the development of AI-driven agricultural decision support systems. |
| 發表成果與AI計畫研究主題相關性 | 本成果運用 UAV 多光譜影像結合機器學習模型，估算小麥田區土壤含水量與產量，展現 AI 於農業影像分析、作物生長監測與早期產量預測之應用，與智慧農業決策支援及 AI 模型建置主題高度相關。 |
