叶绿素含量是植物生长中的重要参数, 与农作物产量密切相关。 无人机遥感技术作为一种新的数据获取手段, 在农业中已得到广泛应用。 以玉米为目标作物, 将具有不同光谱响应函数的两种轻小型多光谱传感器(MCA和Sequoia), 同时搭载在六旋翼无人机上, 获取不同氮肥水平下大田玉米花期的多光谱影像。 利用无人机影像空间分辨率高的特点, 在小区尺度上, 分别计算了基于两种多光谱传感器的各26种植被指数, 并将其与地面实测的叶绿素含量(SPAD)值进行回归分析, 研究不同波段反射率对SPAD值的敏感性, 利用不同多光谱传感器及植被指数预测SPAD值的精度及稳定性。 结果表明, 对于具有较宽波段的Sequoia, 在550 nm(绿波段)、 735 nm(红边波段)的反射率对SPAD值的变化较敏感, 其中, 550 nm与SPAD值的相关系数最大(R2=0.802 9)。 而对于较窄波段的MCA, 720 nm(红边波段)的反射率与SPAD值具有较高的相关性(R2=0.724 8), 550 nm(绿波段)次之。 此外, 由于两传感器红波段的中心波长和波段宽度不同, 660 nm(Sequoia)反射率与SPAD值的相关系数为0.778 6, 而680 nm(MCA)反射率与SPAD值的相关性较小, 仅为0.488 6。 利用无人机多光谱遥感技术预测大田玉米的SPAD值精度较高, 但对于不同的多光谱传感器而言, 同一植被指数却表现出较大的差异, 其中, 红波段和近红外波段组合构造的植被指数RVI, NDVI, PVI和MSR差异较大, 具有较宽波段的Sequoia传感器优于窄波段的MCA; 此外, 对于Sequoia相机, GNDVI与RENDVI预测SPAD值的精度较高, RMSE分别为3.699和3.691; 对于MCA相机, RENDVI预测精度最高(RMSE=3.742), GNDVI预测精度低于RENDVI(RMSE=3.912); 两传感器中MCARI/OSAVI预测SPAD值精度均较低, RMSE分别为7.389(Sequoia)和7.361(MCA)。 在所有的植被指数中, 利用绿波段和近红外波构造的植被指数(G类), 以及用红边波段和近红外波段构造的植被指数(RE类), 预测SPAD值精度更高, 均高于红外和近红外波段构造的植被指数; 利用更多波段(三个及以上)组合构造的复杂植被指数, 并不能显著提高预测精度。 就预测模型而言, MCARI1更适用于对数模型, 可有效提高预测精度, 而其他植被指数变化不显著。 研究还发现, 在小区水平SPAD值的预测方面, 除NDVI和TVI, Sequoia相机对于不同氮肥条件下植被覆盖度、 阴影和裸露土壤等环境背景因素具有较强的抗干扰能力; 而对于MCA相机来说, TVI, DVI, MSAVI2, RDVI和MSAVI对环境背景因素非常敏感, 预测SPAD精度低; 此外, 去除环境背景因素并不总是能够提高SPAD值的预测精度。 本研究对于利用无人机多光谱遥感技术进行高精度的叶绿素含量预测具有指导意义, 对于精准农业的推广和应用具有一定的借鉴价值。

Chlorophyll content is an important parameter in plant growth and is closely related to crop yield. Unmanned aerial vehicle (UVA) remote sensing technology as a new means of data acquisition, has been widely used in agriculture. In this study, take maize as an example, two light and small multispectral sensors (MCA and Sequoia) with different spectral response functions were simultaneously mounted on a six-rotor UAV. Multispectral sensors were used to collect multispectral imagery of maize during the flowering stages under different levels of nitrogen fertilizers. At the plot level, the 26 vegetation indices based on two kinds of multi-spectral sensors were calculated and regressed with the chlorophyll content (called Soil and Plant Analyzer Development (SPAD) values) measured on the ground. The sensitivity of different band reflectivity to SPAD value were analyzed. And the accuracy and stability of SPAD values predicted by vegetation indicesbased on two multi-spectral sensors were also analyzed. The results showed that for the broadband Sequoia, the reflectance of 550 nm (green band) and 735 nm (red-edge band) is more sensitive to the change of the SPAD values, and the correlation coefficient of 550 nm and SPAD values is the largest (R2=0.802 9). For the narrowband MCA, the reflectance of 720 nm (red-edge band) has high correlation with SPAD value (R2=0.724 8), followed by the 550 nm. In addition, the correlation coefficient between the reflectance of 660 nm (Sequoia) and the SPAD value is 0.778 6, and the correlation coefficient between the reflectance of 680 (MCA) and the SPAD value is 0.488 6, which may because of the difference of central wavelength and the wavelength width. Using multi-spectral remote sensing technology of UAV to predict the SPAD values of field maize had a high accuracy, but the same vegetation index showed a great difference for different multi-spectral sensors. Among them, there were significant difference in RVI, DNVI, PVI and MSR. The broadband Sequoia is superior to the narrowband MCA. In addition, for sequoia camera, the GNDVI and RENDVI predicted the SPAD value with high accuracy, RMSE is 3.699 and 3.691, respectively. For MCA camera, RENDVI had the highest prediction accuracy (RMSE=3.742), followed by the GNDVI (RMSE=3.912). The MCARI/OSAVI with lower accuracy, the RMSE is 7.389 (Sequoia) and 7.361 (MCA). In all of the vegetation indices, the vegetation indices that using green, NIR bands and the vegetation indices constructed with red and near infrared bands were used to predict the SPAD values more accurate, which were higher than the vegetation index constructed in the red and near infrared bands. The use of complex vegetation indices constructed with more bands (three or more) did not significantly improve the prediction accuracy. For the prediction model, MCARI1 was more suitable for logarithm Model, which can effectively improve the prediction accuracy. The study also found that prediction the SPAD values in the plot level, Sequoia cameras have strong anti-jamming capability for environmental factors such as vegetation coverage, shadows and exposed soil, except for NDVI and TVI. For MCA cameras, TVI, DVI, MSAVI2, RDVI and MSAVI were very sensitive to environmental background and with a low accuracy of SPAD prediction. In addition, removal of environmental background did not always improve predictive accuracy of SPAD. This study is instructive for the prediction of high-accuracy chlorophyll content using UAV multispectral remote sensing technology, and has certain reference value for the popularization and application of precision agriculture.