Multi-Spatial Vertical Scale Leaf Chlorophyll Content Monitoring of Soybean Based on UAV Multi-Spectral Remote Sensing

基於無人機多光譜遙感的多空間垂直尺度大豆葉片葉綠素含量監測

10.6937/TWC.202406_72(2).0003

ZIJUN TANG(唐子竣)；YOUZHEN XIANG(向友珍)；HONGZHAO SHI(史鴻桌)；DONGMEI LI(李冬梅)

Soybean ； Multi-spectral ； Vertical scale ； Vegetation index ； Machine learning ； 大豆 ； 多光譜 ； 垂直尺度 ； 植被指數 ； 機器學習

台灣水利

72卷2期（2024 / 06 / 01）

50 - 65

英文;繁體中文

Using remote sensing technology to quickly and non-destructively predict the chlorophyll content of soybean leaves is very important for field management. In this study, we employed unmanned aerial vehicle multi-spectral imaging technology to collect spectral data during the flowering period of soybean in 2022 and 2023. Six-teen empirical vegetation indices related to crop physiological growth indicators were selected. Vegetation indices significantly correlated (P < 0.05) with soybean leaf SPAD values at each vertical layer were then screened and integrated as input variables for the models. Prediction models for soybean SPAD values at different vertical scales were established using Partial Least Squares Regression (PLSR), Random Forest (RF), and Support Vector Machine (SVM) models, and the accuracy of the models was validated. The results demonstrated significant correlations (P < 0.05) between various indices and the SPAD values of soybean leaves in different layers. The SPAD values of soybean leaves decreased gradually from top to bottom, with the upper layer showing the highest average correlation coefficient with the vegetation indices, around 6.6% to 9.9% higher than the other leaf positions. Encouragingly, the RF model provided the best fit for estimating the SPAD values of soybean leaves, particularly for the upper layer. The RF model achieved a high R2 of 0.820, with an RMSE of 2.268 and an MRE of 4.434% for the validation set. The results of this study offer technical backing for the remote sensing assessment of soybean chlorophyll content across various spatial scales.

利用遙感技術快速、無損地預測大豆葉片葉綠素含量對田間管理至關重要。本研究利用無人機多光譜成像技術，采集2022年和2023年大豆開花期的多光譜影像數據。選取了16個與作物生理生長指標相關的經驗植被指數。然後篩選出與大豆葉片SPAD值在各垂直層次上顯著相關的植被指數（P<0.05），並將其作為模型的輸入變量。利用偏最小二乘回歸（Partial Least Squares Regression, PLSR）、隨機森林（Random Forest, RF）和支持向量機（Support Vector Machine, SVM）模型建立了不同垂直尺度下大豆SPAD值的預測模型，並對模型的精度進行了驗證。結果表明，各指標與大豆葉片不同層次的SPAD值之間均存在顯著的相關性（P<0.05）。大豆葉片SPAD值自上而下逐漸降低，上層與植被指數的平均相關系數最高，比其他葉位高6.6%～9.9%。令人鼓舞的是，RF模型對大豆葉片SPAD值的估算效果最好，尤其是上層。RF模型R^2為0.820，RMSE為2.268，驗證集的MRE為4.434%。本研究結果為不同空間尺度大豆葉綠素含量的遙感估測提供了技術支撐。