基于光谱的土壤氮含量预测模型泛化能力弱是制约其推广应用的瓶颈。 鉴于特征提取及非线性表达能力方面的优势, 深度学习模型具有较强的泛化能力。 提出一种融合自动编码器和卷积神经网络(Encoder-CNN)的土壤氮含量光谱预测模型, 探索模型结构和参数对模型性能的影响。 根据以往研究成果和相关性分析, 获得180个与氮含量强相关的波长, 将其作为Encoder-CNN模型输入, 而将土壤氮含量作为模型输出。 Encoder-CNN模型利用自动编码器的编码部分进行光谱数据降维, 然后输入到卷积神经网络进行土壤氮含量预测。 设计2种网络结构, 每种网络结构包含2种不同参数设置, 共4个模型, 用以探索Encoder-CNN土壤氮含量光谱预测模型结构和参数对模型性能的影响。 利用公开数据集LUCAS对模型进行训练。 按3σ原则对公开数据集LUCAS进行异常值检测与处理, 获得20 791个数据, 其中18 711个样本作为训练集, 2 080个样本作为测试集, 对Encoder-CNN模型进行训练。 结果表明: 对于自动编码器, 在相同隐含层数下, 最后的隐含层神经元个数为30时, 复现效果最优。 增加隐含层数, 会提升复现效果。 增加卷积核数量, 特别是尺寸为1×1卷积核, 能够提高模型的预测性能与可靠性。 增加池化层的网络结构, 模型预测精度提升至0.90以上。 增加全连接层神经元数量也会提升模型性能。 利用自采集的黑龙江黑土实时光谱数据集进行模型迁移, 观察模型泛化能力。 当模型迭代100次后, 在黑龙江数据集上的预测精度即可达到0.90以上; 当迭代次数为900时, 模型在训练集和测试集上的预测精度可以达到0.98。 结果表明, 所构建的Encoder-CNN土壤氮含量光谱预测模型具有较好的泛化能力。

Poor generalization ability of soil nitrogen content prediction models based on spectroscopy is the bottleneck of its actual application in agriculture production. However, the deep learning model shows strong potential for generalization because of its automatic feature extraction and excellent nonlinear expression. In this paper, a spectral prediction model of soil nitrogen content combining the auto-encoder and convolutional neural network (Encoder-CNN) was proposed, the influence of model structure and parameters on model performance was explored, and its generalization ability was investigated. After researching the previous references and analyzing the correlation between wavelengths and soil nitrogen content, 180 wavelengths with the highest correlation were selected and taken as the input of the Encoder-CNN model. The output of the Encoder-CNN model was the soil nitrogen content. The Encoder-CNN model first used the auto-encoder to reduce the dimension of 180 wavelengths and then predicted the soil nitrogen content by its convolutional neural network. Two network structures were designed. Each network structure had two different parameter settings. Therefore, four models were used to explore the effects of structure and parameters of the Encoder-CNN soil nitrogen content spectral prediction model on modeling performance. Encoder CNN models were trained by the LUCAS data set. According to the 3 σ principle, 20 791 data were obtained from LUCAS and then divided into a training set (18 711) and test set (2 080). The results were analyzed and discussed, and several conclusions were achieved in this research. The reproduction effect of the automatic encoder reached the best when the number of neurons in the last hidden layer was set to 30 with the same number of hidden layers as the others; the more hidden layers, the better the reproduction effect. As for the prediction part based on CNN, increasing the number of convolution kernels, especially 1×1 convolution kernels, could improve the prediction performance and reliability. By adding pooling layer in CNN, the model's prediction accuracy was improved to more than 0.90. The model's performance could also be improved by increasing the number of neurons in the full junction layer. The Encoder-CNN model built by the LUCAS data set was migrated to the Heilongjiang data set, and the generalization ability of the model was observed and evaluated. The prediction accuracy of the model, which was trained 100 times by the Heilongjiang data set, could reach more than 0.90. When the number of iterations was set to 900, the model's prediction accuracy could be as high as 0.98. The results showed that the proposed Encoder-CNN spectral prediction model of soil nitrogen content had good generalization ability, and it could be used to detect soil nitrogen content after the model migration process.