图1 核桃果实不同生育时期叶片含水率 对照、W1、W2、W3、W4灌水定额分别为9 000、5 250、6 000、7 500、8 250 m3 · hm-2

Fig. 1 Leaf moisture content of walnuts fruits at different growth stages The irrigation quotas for the control，W1，W2，W3，and W4 are 9 000，5 250，6 000，7 500，and 8 250 m3 · hm-2，respectively

图2 叶片异常值的剔除

Fig. 2 Removal of blade outliers

图3 21种原始光谱数据预处理转换与叶片含水量的相关系 各预处理方法详见表2

Fig. 3 Correlation coefficient between 21 original spectral preprocessing methods and leaf moisture content Each preprocessing method is detailed in Table 2

图4 ‘温185’核桃叶片不同预处理后的光谱数据 A：原始光谱图（RAW）；B：一阶导数预处理（FD）；C：去趋势预处理（Detrend）；D：卷积平滑 + 一阶导数组合预处理（SG + FD）；E：一阶导数 + 标准正态变化组合预处理（SNV + FD）；F：标准正态变化 + 去趋势组合预处理（SNV + Detrend）

Fig. 4 Spectral data of walnut leaves with different pretreatments at‘Wen 185’walnut A：Raw spectrogram；B：First derivative pre-processing（FD）；C：Detrend preprocessing；D：Convolutional smoothing combined with first-order derivative preprocessing（SG + FD）；E：First derivative and standard normal variation combination preprocessing （SNV + FD）；F：Standard normal variation and detrend combination preprocessing（SNV + Detrend）

图5 不同预处理CARS算法的波长变量筛选过程 a：变量数量变化趋势；b：RMSECV值变化趋势；c：每个变量回归系数值变化趋势

Fig. 5 The wavelength variable selection process of different preprocessing CARS algorithms a：Trend of changes in the number of variables；b：RMSECV value change trend；c：The trend of regression coefficient values for each variable

图6 不同预处理CARS算法的提取特征结果

Fig. 6 Feature extraction results of different preprocessing CARS algorithms

图7 不同预处理UVE提取特征波长原理 A、C、E、G、I、K：各预处理波段筛选过程，通过加入白噪声变量，根据PLS模型交叉留一法得到各变量回归系数，2条水平线分别为UVE变量筛选出的上下阈值线；B、D、F、H、J、L：各预处理提取的特征波段筛选结果

Fig. 7 Principle of extracting feature wavelength from different preprocessed UVE A，C，E，G，I，K：During the screening process of each preprocessing band，the regression coefficients of each variable are obtained by adding white noise variables and using the PLS model cross leave one method，Two horizontal lines represent the upper and lower threshold lines filtered by UVE variables. B，D，F，H，J，L：Screening results of feature bands extracted by each preprocessing

图8 不同预处理SPA提取特征波长结果 A、C、E、G、I、K为各预处理波段筛选过程，代表SPA多元线性回归模型选择的变量数图，红色方块代表的变量数，此时的预测平均标准偏差最小；B、D、F、H、J、L为各预处理波段筛选结果，其中红色方块标记的是通过SPA（连续投影算法）提取的特征波段

Fig. 8 Feature wavelength results extracted from different preprocessed SPA A，C，E，G，I，and K are the screening processes for each preprocessing band，representing the number of variables selected by the SPA multiple linear regression model. The red square represents the number of variables，and the average standard deviation of the prediction is the smallest at this time；B，D，F，H，J，and L are the filtering results of each preprocessing band，where the red square marks the feature bands extracted through SPA（Continuous Projection Algorithm）

图9 不同预处理BPSO提取特征波长原理 A、C、E、G、I、K为各预处理波段筛选过程；B、D、F、H、J、L为各预处理波段筛选结果；适应度曲线变化和筛选误差成正比，随着迭代次数的增加，适应度曲线呈下降趋势，模型误差也在减小，当误差下降到最低值时，筛选出的特征波长即为最优特征波长

Fig. 9 Principle of extracting feature wavelength from different preprocessed BPSO A，C，E，G，I，K are the screening processes for each preprocessing band；B，D，F，H，J，L are the screening results of each preprocessing band； The change in fitness curve is directly proportional to the screening error. As the number of iterations increases，the fitness curve shows a downward trend，and the model error also decreases. When the error drops to the lowest value，the selected feature wavelength is the optimal feature wavelength

图10 基于FD预处理经4种特征波段提取构建不同模型预测结果

Fig. 10 Prediction results of different models constructed based on FD preprocessing and extraction of four feature bands