在拔节期分4次采集了6个施氮水平下的冬小麦冠层图像, 同步进行取样并以凯氏定氮法测定叶片含氮量, 进而计算叶片氮积累量。 利用随机森林算法分割冠层图像之后提取冠层覆盖度、 可见光波段(R, G和B)三个分量及其衍生的5个色彩指数。 以冠层覆盖度外加色彩指数、 色彩分量的两种非线性回归, 以及人工神经网络、 支持向量回归、 随机森林3种机器学习算法建立了冬小麦叶片氮积累量的估算模型。 结果表明利用色彩指数的非线性回归模型的估算精度稍低于其他方法, 而随机森林算法的拟合精度最高, 但存在明显的过拟合现象。 其他三种方法, 即以冠层覆盖度及色彩分量为输入变量的非线性回归、 支持向量回归和人工神经网络方法, 均具有较高的拟合精度和泛化性能。

In order to study the feasibility of using digital image analysis and machine learning algorithm to estimate leaf nitrogen accumulation (LNA) of winter wheat at canopy level, digital images of winter wheat canopies grown under six levels of nitrogen application rate were taken for four times during the elongation stage. Meanwhile, wheat plants were sampled to measure LNA. The random forest method using CIEL*a*b* components was used to segment wheat plant from soil background and then extract canopy cover, RGB components of sRGB color space and compute five color indices derived from RGB components. Correlation analysis was carried out to identify the relationship between LNA and canopy cover (CC), RGB components, and five color indices. Two kinds of nonlinear least squares regression models (NLS) with different independent variables of color components and color indices, and three machine learning algorithmic of artificial neural network (ANN), support vector regression (SVR), and random forests method (RF) were used to estimate winter wheat leaf nitrogen accumulation. All three machine learning algorithm had four input variables of CC, R, G, and B. The results showed that, CC, R and G component of sRGB color space, and five color indices derived from RGB components showed significant correlations with LNA during the elongation stage. CC revealed the highest correlation with LNA. The lowest accuracy in estimation LNA was achieved by using nonlinear least square model with CC and color indices, and RF had showed the problem of overfitting. The other three methods of LNA with CC and RGB components, ANN, and SVR had showed good performance with higher R2 (0.851, 0.845, and 0.862) and lower RMSE (19.440, 19.820, and 18.698) for model calibration and validation, revealing good generalization ability.