土壤pH值是影响土壤养分转化和土壤肥力的关键因素, 使用近红外光谱技术对土壤pH值进行检测可为土壤资源的开发利用提供重要依据。 卷积神经网络作为深度学习在人工智能方面的典型算法, 由于其结构具备“局部感知, 权值共享”的能力, 因此不仅能够对复杂的光谱数据进行特征抽取, 还能够减少网络的训练参数, 提高网络的运算效率。 将卷积神经网络用于近红外光谱的建模分析, 并提出一种基于一维卷积的卷积神经网络和近红外光谱的土壤pH值预测方法。 网络由Python语言调用Tensorflow工具包搭建而成, 其结构由输入层、 卷积层、 池化层以及全连接层四部分组成。 以欧洲统计局在2008年—2012年开展的土地利用及覆盖面积统计调查所收集的矿物质土壤光谱样本数据集为研究对象, 为消除光谱中存在的基线漂移, 提高信噪比, 对原始可见光近红外光谱(400～2 500 nm)进行一阶导数和Savitzky-Golay平滑处理。 在模型训练过程中, 随机选取15 000个样本作为训练集, 剩余的2 272个样本作为测试集, 探讨不同的卷积层个数及训练迭代次数对模型性能的影响, 并采用ReLU激活函数及Adam优化器防止模型出现梯度消失现象, 提高模型的稳定性, 之后通过分析模型的拟合优度和运算成本确定模型的最佳性能, 最后将网络模型与传统的BP和PLSR模型进行对比。 结果显示, 当模型迭代次数为2 500次, 卷积层个数为4层时, 模型达到最佳状态, 模型对训练集的均方误差从1.898降到了0.097; 模型对测试集的拟合优度为0.909, 分别比BP和PLSR模型高0.117和0.218。 使用卷积神经网络可以对土壤近红外光谱的内部特征信息进行抽取, 从而实现对大面积土壤pH值的高效准确预测。 CNNR模型可对农作物的合理栽种及精准施肥提供指导, 从而达到土壤结构稳定和可持续发展的目的。 基于卷积神经网络的近红外光谱回归方法也可以推广到其他土壤信息研究。

The soil pH is the key factor affecting the transformation of soil nutrients and the soil fertility. The detection of pH value of soil by near-infrared spectroscopy can provide an important basis for the development and utilization of soil resources. As a typical algorithm of deep learning in artificial intelligence, the convolutional neural network can not only extract the characteristics of complex spectral data but also reduce the training parameters of the network and improve the efficiency of network operation, because its structure has the ability of “local perception, weight sharing”. In this paper, the convolution neural network is applied to the modeling and analysis of the near-infrared spectrum, and a soil pH prediction method based on convolution neural network and the near-infrared spectrum is proposed. The network is built by Python calling Tensor Flow toolkit, and its structure is composed of the input layer, convolution layer, pooling layer and full connection layer. The spectral sample dataset of mineral soils, collected from the Statistical Survey of Land Use and Coverage conducted by the European Statistical Office in 2008—2012, was employed as an object of study. In order to eliminate the baseline drift in the spectrum and improve the signal-to-noise ratio, the first derivative and Savitzky-Golay smoothing of the original visible near-infrared spectrum (400~2 500 nm) were carried out. In the model training process, 15 000 samples are randomly selected as the training set, and the remaining 2 272 samples are selected as the test set. The effects of the number of convolution layers and training iterations on the model performance are discussed. The ReLU activation function and Adam optimizer are used to prevent the gradient disappearance of the model and improve the stability of the model. Then, the goodness of fit of the model is analyzed and calculated, and finally, the network model is compared with the traditional BP and PLSR models. The experimental results show that when the number of iterations of the model is 2 500, and the number of convolution layers is 4, the model reaches the best performance, and the mean square error of the training set is reduced from 1.898 to 0.097; the goodness of fit of the test set is 0.909, which is 0.117 and 0.218 higher than BP and PLSR models respectively. The results indicate that convolution neural network can extract the internal characteristic information of soil near-infrared spectrum, so as to realize efficient and accurate prediction of soil pH on a large scale. CNNR model can provide guidance for crop planting and precision fertilization to achieve the goal of soil structure stability and sustainable development. The convolution neural network-based NIRS regression method can also be applied to other soil information research.