利用無人機擷取之多光譜影像近似燕麥芻料性狀

Abstract

燕麥（Avena sativa L.）是重要的穀類作物，為單年生禾本科植物，除了穀粒供人類食用外，也常用作動物芻料。育成優良的芻料燕麥品種除關注產量外，還需評估其營養價值 (Nutritive values, NV)，然一般產量性狀與營養價值之量測方法耗時費力，不利於育種的進行。透過無人機（Unmanned aerial vehicle, UAV）搭載多光譜相機，獲取可反映植物生長狀態之特定波長的反射光譜影像，近似產量性狀和營養價值為可能的解方。因此，本研究的目的為利用無人機擷取之多光譜影像近似燕麥芻料性狀。本研究於國立臺灣大學附設農業試驗場進行為期兩年的田間試驗，播種時間分別為2020年10月28日和2021年11月3日。第一年試驗種植4個具有不同特性的燕麥品系，第二年除第一年的4個品系，另增加8個，共有12個燕麥品系。第一年在其中一個品系開始抽穗時進行收穫，於收穫當日利用無人機搭載多光譜相機拍攝影像，共有7個日期的多光譜影像；第二年則於植株達抽穗期時進行多光譜影像拍攝及收穫。本研究利用五個波段之反射光譜值及由反射光譜衍生之植生指數 (Vegetation index, VI) 作為自變數，以性狀數據做為應變數，並以最佳子集迴歸 (Best subsets regression) 及交叉驗證 (Cross-validation) 建立最佳模型。結果顯示，波段模型和植生指數模型的表現相似，波段模型對於營養價值的預測性較好，其中，對Ash的預測效果最差，預測值與實際觀測值之皮爾森相關係數 (Pearson’s correlation coefficient) 僅0.20，其他營養價值的相關係數則介於0.39到0.54；植生指數模型對於產量性狀的預測性較好，FWY和DWY的相關係數分別為0.70和0.65。本研究建立了一套非破壞性且高通量的芻料性狀分析方法，利用無人機拍攝多光譜影像即可快速估算芻料燕麥的產量及營養價值。未來的研究可擴大試驗範圍，包括更多的收穫期、品系及地點，以增加數據的多樣性，從而使模型能更廣泛地應用於芻料燕麥的育種。

Oats (Avena sativa L.) serve as both human food and animal feed. Breeding high-quality forage oat varieties not only needs to consider the yield but also requires evaluating their nutritive values (NV). Traditional methods for measuring yield traits and NV are time-consuming and labor-intensive. The present study aims to approximate oat forage traits using multispectral images captured by unmanned aerial vehicle (UAV). Field experiments were conducted over two years at the Experimental Farm of National Taiwan University. Oats were sowed on October 28, 2020 and November 3, 2021. In the first year, 4 oat lines with different characteristics were sowed, while 8 more lines were included in the second-year experiment. Reflectance of five wavelengths, as well as eight vegetation indices (VI) calculated based on the reflectance data, were used for trait prediction model fitting using best subsets regression and cross-validation. Results showed similar prediction ability between the direct use of reflectance and VI. The reflectance models were better at predicting NVs, with the poorest prediction for Ash, having a Pearson’s correlation coefficient of only 0.20 between the predicted and actual values. For other NVs, the correlation coefficients ranged from 0.39 to 0.54. The VI models were better at predicting yield traits, with correlation coefficients for fresh weight yield (FWY) and dry weight yield (DWY) being 0.70 and 0.65, respectively. This study established a non-destructive, high-throughput method for analyzing oat forage traits using UAV-captured multispectral images, with potential applications in forage oat breeding. The experiment can be extended in the future to include additional harvest periods, oat lines, and locations. This expansion would enhance data diversity and extend the model's predictive ability.