提出一种改进自动编码器方法, 用于利用近红外光谱预测大尺度下土壤有机质含量等级。首先, 提出改进自动编码器算法框架, 将传统的用于重建输出的自动编码器与分类器相结合; 对改进自动编码器中的损失函数进行定义。然后, 将改进自动编码器应用于预测土壤有机质含量等级的近红外光谱分析建模问题中, 使用双层前馈神经网络实现了改进自动编码器的编码器、解码器和分类器。最后, 使用大尺度土壤光谱数据集对模型进行训练, 预测土壤有机质含量等级, 并与主成分回归、支持向量机等方法的效果进行对比。实验结果表明, 基于改进自动编码器的土壤有机质含量等级分类准确率为63.05%, 高于其他方法。利用该模型预测大尺度下土壤有机质含量等级有较好的表现。

This paper presents a calibration model, namely, improved auto encoder, which can be used to predict the grade of soil organic matter content in large scale based on near infrared spectroscopy. First, the framework of improved auto encoder model was proposed, which combined traditional auto encoder and classifier, and loss function of the model is defined. Then, the proposed improved auto encoder was applied to predict the grade of soil organic matter content based on near infrared spectroscopy. The encoder, decoder and classifier were implemented with two-layer feed-forward neural networks. Finally, a large scale soil spectral dataset was used to train the model for predicting the grade of soil organic matter content. The performance was compared with the results of principal component regression and support vector machine. The results show that the classification accuracy of soil organic matter content grades based on the proposed improved auto encoder model is 63.05%, which is better than other methods. This model can be used to predict the grades of soil organic matter content in a large scale.