Abstract Leaf chlorophyll content (LCC) is a crucial indicator of nutrition in crop plants and can be applied to assess the adequacy of nitrogen (N) fertilizer for crops while reducing N losses to farmland. This study estimated the LCC of maize and wheat, and comprehensively examined the effects of the spectral information and spatial scale of unmanned aerial vehicle (UAV) imagery, and the effects of phenotype and phenology on LCC estimation. A Cubert S185 hyperspectral camera onboard a DJI M600 Pro was used to conduct six flight missions over a long-term experimental field with five N applications (0, 70, 140, 210, and 280 kg N ha−1) and two irrigation levels (60% and 80% field water capacity) during the growing seasons of wheat and maize in 2019. Four regression algorithms, that is, multi-variable linear regression, random forest, backpropagation neural network, and support vector machine, were used for modeling. Leaf, canopy, and hybrid scale hyperspectral variables (H-variables) were used as inputs for the statistical LCC models. Optimal H-variables for modeling were determined by Pearson correlation filtering followed by a recursive feature elimination procedure. The results showed that (1) H-variables at the canopy- and leaf-scales were appropriate for wheat and maize LCC estimation, respectively; (2) the robustness of LCC estimation was in the order of the flowering stage > heading stage > grain filling stage for wheat and early grain filling stage > flowering stage > jointing stage for maize; (3) the reflectance of the red edge, green, and blue bands were the most important inputs for LCC modeling, and the optimal vegetation indices differed for the various growth stages and crops; and (4) all four algorithms maintained an acceptable accuracy with respect to LCC estimation, although random forest and support vector machine were slightly better. This study is valuable for the design of appropriate schemes for the spectral and scale issues of UAV sensors for LCC estimation regarding specific crop phenotype and phenology periods, and further boosts the applications of UAVs in precision agriculture.

中文翻译：

---

通过多尺度最优无人机高光谱数据估算作物叶片叶绿素含量

摘要 叶绿素含量（LCC）是衡量作物营养的重要指标，可用于评价作物施氮（N）的充足性，同时减少农田氮素流失。本研究估计了玉米和小麦的 LCC，并综合考察了无人机 (UAV) 图像的光谱信息和空间尺度的影响，以及表型和物候对 LCC 估计的影响。DJI M600 Pro 搭载的 Cubert S185 高光谱相机用于在长期试验田执行六次飞行任务，使用五次氮应用（0、70、140、210 和 280 kg N ha−1）和两个灌溉水平（ 60% 和 80% 的田间持水量）在 2019 年小麦和玉米的生长季节。四种回归算法，即多变量线性回归、随机森林、反向传播神经网络和支持向量机被用于建模。叶、冠层和混合尺度高光谱变量（H 变量）用作统计 LCC 模型的输入。用于建模的最佳 H 变量由 Pearson 相关过滤和递归特征消除程序确定。结果表明：（1）冠层和叶尺度的H变量分别适用于小麦和玉米的LCC估计；(2)LCC估计的稳健性顺序为小麦开花期>抽穗期>灌浆期，玉米灌浆期早期>开花期>拔节期；(3) 红色边缘、绿色和蓝色波段的反射率是 LCC 建模最重要的输入，不同生长阶段和作物的最佳植被指数不同；(4) 尽管随机森林和支持向量机稍好一些，但所有四种算法在 LCC 估计方面都保持了可接受的精度。该研究对于针对特定作物表型和物候期的 LCC 估计无人机传感器的光谱和尺度问题设计适当的方案具有价值，并进一步推动无人机在精准农业中的应用。