基于双分支卷积网络的玉米叶片叶绿素含量高光谱和多光谱协同反演

针对智慧农业中叶绿素的精准预测问题，本文提出了基于双分支网络的玉米叶片叶绿素含量高光谱与多光谱协同反演的方法。使用欠完备自编码器进行数据降维，捕捉数据中最为显著的特征，使降维后的数据可以代替原始数据进行训练，从而加快训练效率，使用双分支卷积网络将多光谱数据用于填充高光谱数据信息，充分利用高光谱数据的空间细节信息，再结合1DCNN建立玉米叶片叶绿素含量预测模型。结果表明，与传统降维算法相比较，欠完备自编码器处理后预测结果最佳，决定系数R2为0.988，均方根误差（RMSE）为0.273，表明使用欠完备自编码器进行降维可以有效提高数据反演精度；与单一的高光谱数据反演模型和多光谱数据反演模型相比，双分支卷积网络预测模型均取得较优的预测结果，R2在0.932以上，RMSE均在1.765以下，表明基于双分支卷积网络的高光谱与多光谱图像协同反演模型可以有效地利用数据的特征；对于其他数据结合本文提及的双分支卷积网络模型进行反演，其R2均在0.905以上，RMSE均在2.149以下，表明该预测模型具有一定的普适性。

Hyperspectral and Multispectral Co-inversion of Chlorophyll Content in Maize Leaves Based on Two-branch Convolutional Network

Aiming at the problem of accurate chlorophyll prediction in smart agriculture, a method of hyperspectral and multispectral synergistic inversion of chlorophyll content in maize leaves was proposed based on two-branch network. The undercomplete self-encoder was used for data dimensionality reduction to capture the most significant features in the data, so that the dimensionality reduced data can be trained instead of the original data to accelerate the training efficiency, and the two-branch convolutional network was used to fill the hyperspectral data with multispectral data to make full use of the spatial detail information of the hyperspectral data, and then combined with the 1DCNN to establish a prediction model of chlorophyll content in maize leaves. The results showed that compared with the traditional dimensionality reduction algorithm, the undercomplete self-encoder processed the best prediction results, with a coefficient of determination R2 of 0.988 and a root mean square error (RMSE) of 0.273, indicating that dimensionality reduction using the undercomplete self-encoder was effective in improving the accuracy of data inversion. Compared with the single hyperspectral data inversion model and the multispectral data inversion model, the two-branch convolutional network prediction models both achieved better prediction results, with R2 above 0.932 and RMSE below 1.765, indicating that the collaborative hyperspectral and multispectral image inversion model based on the two-branch convolutional network can make effective use of the features of the data. For the other data combined with the mentioned two-branch convolutional network model for the inverse model, the R2 was above 0.905 and the RMSE was below 2.149, which indicated that the prediction model had a certain degree of universality.