叶面积指数（LAI）是评价作物长势的重要参数, 快速、 准确、 低成本地获取作物LAI对于指导作物田间管理有重要的意义。 为了低成本获取多种作物的LAI, 基于多源信息和深度学习构建了通用的LAI预测模型。 在大豆、 小麦、 花生、 玉米四种作物的六个生长时期进行了大田实验, 以获取用于建模的多源信息。 使用航拍无人机获取作物低空可见光图像、 红边图像和近红外图像等多光谱图像信息, 此外还采集相关的一维数据信息, 包括无人机飞行姿态、 拍摄高度、 作物生长状态和环境光照。 借助深度学习出色的图像和数据处理能力建立基于复杂输入信息的LAI预测模型, 考虑到一维数据也要参与模型的训练过程, 在设计模型时, 采用了组合型网络架构。 在卷积神经网络（CNN）算法提取图像深度特征的基础上加入了LightGBM算法用于结合图像特征和一维数据实现作物LAI的最终预测。 CNN模型部分使用了VGG19, ResNet50, Inception V3和DenseNet201四种常见的结构。 为了更好地说明CNN模型提取图像特征的能力, 分析了不同图像输入下四种模型的作物分类情况。 结果表明, 以可见光、 红边和近红外图像为输入时, 四种模型的分类准确度均相较于仅有可见光图像时有所提高, 尤其是基于Inception V3和DenseNet201的两种模型分类准确率均达到99%以上, 证明了CNN模型提取多光谱图像特征的有效性。 将图像特征作为LightGBM模型的输入信息预测LAI时, 实测值与预测值的R2最大为0.819 2, 而在输入中加入一维数据信息后, 模型的R2均可达到0.9以上, 说明多源信息输入对于提高LAI预测模型的准确度有重要作用。 该研究建立的模型可以针对不同的作物进行LAI的预测, 不需要对多光谱图像进行复杂的处理, 因此, 该研究可以实现LAI的低成本、 快速预测, 同时可以获得较高的预测准确度。

Leaf area index (LAI) is an important parameter for evaluating crop growth, rapid, accurate and low-cost acquisition of LAI has great significance for guiding crop field management. To achieve low-cost acquisition of LAI for multiple crops, the general LAI prediction models were built based on multi-source information and deep learning. The field experiments were carried out in six growth periods of soybean, wheat, peanut, and maize to obtain multi-source information for modeling. In addition, relevant one-dimensional data were collected, including UAV flight attitude angles, image capture height, crop growth states and environmental illumination. With the help of the excellent image and data processing ability of deep learning, the LAI prediction models were built accurately based on complex input information. Considering that the one-dimensional data also participated in the training process of the models, the combined network architecture was adopted in the design of the models. Based on extracting image depth features by the convolutional neural network (CNN) algorithm, the LightGBM (Light Gradient Boosting Machine Method) algorithm was added to realize the final prediction of crop LAI by combining image features and one-dimensional data. Four common network structures, VGG19, ResNet50, Inception V3 and DenseNet201, were used in four CNN models. In order to better illustrate the ability of CNN models to extract image features, the crop classification results of the four models under different image inputs were analyzed. The results showed that the classification accuracyof the four models with inputs using multispectral images was better than that of inputs using visible images only. The classification accuracy of the models based on Inception V3 and DenseNet201 was more than 99%, which proved the validity of the CNN model in extracting multispectral image features. Taking the image features as inputs of the LightGBM model to predict LAI, the results were shownthat the maximum R2 betweenthe measured value and the predicted value of LAI is 0.819 2. After one-dimensional data were addedtothe inputs, the R2 of the models can reach more than 0.9, which indicates that multi-source information inputs play an important role in improving the accuracy of the LAI prediction models. The models established in this study can predict LAI for multiple crops without the complex processing for multispectral images. Therefore, this study can realize the low-cost and rapid prediction of LAI and have high LAI prediction accuracy.