基于1D-CNN的土壤全氮近红外光谱预测模型

摘要：【目的】传统的基于近红外光谱数据预测土壤全氮的方法需要对原始光谱数据做复杂的预处理，筛选出与土壤全氮含量相关性高的敏感波长之后进行模型的回归拟合。本文提出一种一维卷积神经网络（1D-CNN）模型，可以在对数据进行简单预处理甚至无处理的情况下达到非常理想的结果，实现用近红外光谱技术对土壤全氮含量的预测。【方法】于江苏无锡采集410个土壤样品，利用半微量开氏法(NY/T 53-1987)测定土壤的全氮含量，并利用NIR Quest 512光谱仪，在室内环境下对每份土壤样品做光谱检测，并用均值中心化(CT)、标准正态变换(SNV)、趋势校正(DT)对光谱进行预处理，运用偏最小二乘回归（PLS）、BP神经网络、1D-CNN方法建立土壤全氮含量的回归预测模型。每种模型在采用不同预处理方法的数据集上做十折交叉验证，记录预测模型的决定系数(R2)和均方根误差(RMSE)的平均值，并对比三种预处理方法对模型精度的影响。【结果】证明了本文提出的1D-CNN模型基于土壤近红外光谱数据预测土壤全氮含量的可靠性。使用原始数据与经均值中心化、标准正态变换、趋势校正预处理的数据训练得到的1D-CNN模型的决定系数分别为0.907、0.931、0.922、0.964，构建的PLS回归模型决定系数为0.856、0.863、0.861、0.880，训练的BP神经网络的决定系数为0.874、0.907、0.901、0.911。【结论】本文提出的1D-CNN模型在原始数据和经预处理的光谱数据上的表现都优于PLS和BP神经网络，且可以证明，对光谱数据进行预处理能够有效提高1D-CNN模型的性能，尤其是趋势校正对模型的提升效果最明显。研究表明，1D-CNN能更好地提取光谱特征并建立其与含氮量的映射关系，有效地避免过拟合，在未经过预处理的光谱数据上依然能够达到一定的精度。

Near-infrared Spectral Prediction Model of Soil Total Nitrogen Based on 1D-CNN

A total of 410 soil samples were collected in Wuxi, Jiangsu, China, and total nitrogen contents. and soil sample spectra were analyzed indoors. The spectral data underwent preprocessing, including mean centering, standard normal variate transformation, and trend correction. Regression prediction models for soil total nitrogen content were established using partial least squares (PLS), back propagation (BP) neural networks, and one-dimensional convolutional neural networks (1D-CNN). Each model underwent ten-fold cross-validation using datasets preprocessed with various methods, and the average values of the coefficient of determination (R²) and root mean square error (RMSE) were recorded to compare the impact of these three preprocessing methods on model accuracy. The results demonstrated the reliability of the 1D-CNN model constructed based on soil near-infrared spectral data. The R² values for the 1D-CNN model trained with raw data and data preprocessed with mean centering, standard normal variate transformation, and trend correction were 0.907, 0.931, 0.922, and 0.964, respectively. In comparison, the R² values for the PLS model were 0.856, 0.863, 0.861, and 0.880, while the BP neural network model''s R² values were 0.874, 0.907, 0.901, and 0.911. The 1D-CNN model consistently outperformed the PLS and BP neural network models on both raw and preprocessed spectral data. Preprocessing the spectral data effectively enhanced the 1D-CNN model''s performance, with trend correction demonstrating the most substantial improvement. Hence, 1D-CNN is adept at extracting spectral features and establishing a robust mapping relationship with nitrogen content, effectively preventing overfitting. Even with unprocessed spectral data, it still achieves a commendable level of accuracy.