基于CARS-CNN的高光谱柑橘叶片含水率可视化研究

柑橘叶片水分亏缺是影响柑橘生长发育的重要因素之一，为研究水分胁迫对柑橘含水率的影响，利用高光谱快速无损检测柑橘叶片含水率，并应用伪彩色处理实现含水率可视化。收集100片柑橘叶片，使用烘干法得到鲜叶和烘干叶片一共500个不同梯度含水率的数据样本，将样本按7∶3的比例划分为训练集（350个样本）和测试集（150个样本），使用决定系数(R2)和均方根误差(RMSE)来评估模型预测的好坏。采用卷积神经网络（CNN）对高光谱数据进行预测，CNN模型使用一维卷积核，一共三层卷积池化层，使用RELU激活函数激活，输出层采用linear激活函数回归预测，使用nadam算法对模型进行优化更新，迭代次数为1 000次；将原始光谱数据和SG，MSC和SNV三种预处理后的光谱数据，与全波段、CARS筛选的特征波段、PCA提取的特征波段组合，导入CNN模型，确定最佳模型为原始光谱数据的CARS-CNN，训练集的R2_c和RMSEC分别为0.967 9和0.016 3，测试集的R2_v和RMSEV分别为0.947 0和0.021 4；原始光谱数据的全波段CNN模型效果其次，训练集的R2_c和RMSEC分别为0.934 3和0.024 9，测试集的R2_v和RMSEV分别为0.915 9和0.028 6。对比了不同预处理方式和特征波长选择的支持向量机回归模型(SVR)、偏最小二乘法回归模型（PLSR）、随机森林模型(RF)组合的最佳结果，将最佳组合模型（原始光谱数据+CARS+PLSR，SNV+PCA+RF，SNV+PCA+SVR）与原始光谱数据的CARS-CNN对比，结果表明，依然是原始光谱数据的CARS-CNN模型预测效果最佳。相较于其他的模型，CARS-CNN模型经过CARS筛选特征波段和卷积核进一步提取特征后，预测精度远高于SVR，PLSR和RF模型。选择训练好的CARS-CNN模型，将高光谱图片导入到模型中，计算每个像素点的含水率，得到伪彩色图像，能够可视化叶片的含水率分布情况。研究结果为柑橘叶片水分含量提供了更快速、更直观、更全面的评估，为研究柑橘叶片水分胁迫提供了依据，为智能灌溉决策的优化提供了参考。

Hyperspectral Visualization of Citrus Leaf Moisture Content Based on CARS-CNN

Water deficit of citrus leaves is one of the important factors affecting the growth of citrus. In order to study the effect of water stress on the moisture content of citrus, hyperspectral technology was used to rapidly and non-destructively detect the moisture content of citrus leaves, and pseudo-color processing was applied to realize the visualization of moisture content. 100 citrus leaves were collected, and 500 leaves with different gradient moisture content were obtained by drying method. The samples were divided into a training set (350 samples) and a testing set (150 samples) according to the ratio of 7∶3. The determination coefficient (R2) and root mean square error (RMSE) was used to evaluate the model’s prediction quality. A convolution neural network (CNN) is used to predict spectrum data. The CNN model uses a one-dimensional convolution kernel with three convolution pooling layers activated by the RELU activation function. The output layer uses a linear activation function for regression prediction, and the nadam algorithm is used to optimize and update the model with 1 000 epochs; The Raw spectrum data and the spectrum data are pretreated by SG, MSC and SNV are used respectively. The full bands, the feature bands screened by CARS and the feature bands extracted by PCA are imported into the CNN model respectively. The best model is CARS-CNN of the Raw spectrum data, the R2_c and RMSEC of the training set are 0.967 9 and 0.016 3 respectively. The R2_v and RMSEV of the testing set are 0.947 0 and 0.021 4, respectively. The effect of the full bands CNN model of the Raw spectrum data is the second, and the R2_c and RMSEC of the training set are 0.934 3 and 0.024 9, respectively. The R2_v and RMSEV of the testing set are 0.915 9 and 0.028 6, respectively; At the same time, the best combined results of the support vector machine regression model (SVR), partial least squares regression model (PLSR) and random forest model (RF) under different pretreatment methods and characteristic wavelength selection were compared. The best combination model (Raw spectrum+CARS+PLSR, SNV+PCA+RF, SNV+PCA+SVR) was compared with CARS-CNN of Raw spectrum data, CARS-CNN model still has the best prediction effect. Compared with other models, the CARS-CNN model has higher prediction accuracy than SVR, PLSR and RF models, after further feature extraction by CARS and convolution kernel. Select the trained CARS-CNN model, import the hyperspectral image into the model, calculate the moisture content of each pixel, and get the pseudo-color image, which can more intuitively display the visual distribution of leaf moisture content. The result provides a faster, more intuitive and more comprehensive assessment of citrus leaf moisture content, a basis for the study of citrus leaf water stress, and a reference for optimising intelligent irrigation decision-making.