土壤粒度对基于近红外离散波长土壤全氮预测精度影响

土壤粒度是对土壤近红外光谱造成严重干扰的主要因素之一。通常在样本前处理阶段采用研磨和过筛土壤来降低土壤粒度干扰，在数据处理阶段通过对连续光谱微分法等数学方法消除土壤粒度干扰。但是对于近红外波段离散波长的建模，至今没有有效的方法消除土壤粒度干扰。为此，提出了土壤粒度修正法以解决土壤粒度干扰消除难题。首先建立土壤粒度修正模型，将农田采集的标准土壤在实验室烘干消除水分后，进行土样配置，得到4个土壤粒度（2.0, 0.9, 0.45, 0.2 mm）和6个全氮浓度等级（0, 0.04, 0.08, 0.12, 0.16, 0.2 g·kg-1）的96个土壤样本。采用MATRIX-Ⅰ型傅里叶变换近红外光谱仪采集土壤样本近红外光谱，计算四个不同粒度(每个粒度包含24个土壤样本)和全部土壤样本在每个波长处（850～2 500 nm）所有样本间吸光度的光谱标准偏差，分析得到土壤粒度的特征波段为1 361和1 870 nm。采用特征波段吸光度比值作为单一输入变量建立SVM土壤粒度分类模型，土壤粒度整体分类准确率为93.8%，表明对土壤粒度进行分类是可行的。选择本研究团队开发的基于近红外波段离散波长(1 070, 1 130, 1 245, 1 375, 1 550, 1 680 nm)吸光度的车载土壤全氮检测仪对提出的土壤粒度修正模型进行验证。结果表明修正后粒度为2.0，0.9和0.45 mm的吸光度和原始土壤吸光度分别降低了62%，74%，111%和61%。表明土壤粒度修正法可以显著减小土壤粒度干扰。最后采用BPNN建立不同吸光度的全氮模型，相较于原始吸光度模型，修正后的土壤吸光度模型R2_v提高了25%。表明提出的土壤粒度修正法可以显著减小土壤粒度对近红外光谱离散波长吸光度的干扰，提高车载土壤全氮检测仪的测量精度。

Effect of Soil Particle Size on Prediction of Soil Total Nitrogen Using Discrete Wavelength NIR Spectral Data

Soil particle size is one recognized factor that cause serious interference to the Near-Infrared (NIR) spectroscopy. Generally, grinding and sieving soil are used to reduce soil particle size interference in the sample pre-processing stage. Mathematical methods such as the continuous spectrum derivative method are used to eliminate soil particle size interference in the data processing stage. However, for the discrete NIR spectral data, so far, there is still no effective methods to eliminate the interference of soil particle size. In this paper, the discrete NIR absorbance data of the soil samples are taken as the research object, to solve the problem of soil particle size interference elimination, and a soil particle size correction method is proposed. Firstly, establishing a soil particle size correction model. After drying the standard soil samples collected from the field to eliminate the interference of soil moisture, the soil samples are prepared. Finally, a total of 96 soil samples under four soil particle sizes (2.0, 0.9, 0.45, 0.2 mm) and six soil total nitrogen (TN) concentration levels (0, 0.04, 0.08, 0.12, 0.16 and 0.2 g·kg-1) were obtained. Calculating the standard deviation of four different particle sizes (each particle size contains 24 soil samples) and all 96 soil samples at each wavelength (850～2 500 nm), and the 1 361 nm and 1 870 nm were confirmed to be the characteristic wavebands of the soil particle size. The characteristic wavebands ratio was used as a single input variable to establish the SVM soil particle size classification model, and the overall classification accuracy of soil particle size was 93.8%. The results showed that it was feasible to use to classify the soil particle size. Based on the above results, a soil particle size correction method is proposed to eliminate the interference of soil particle size to the discrete NIR spectral data. Our team selected the six discrete NIR wavebands (1 070, 1 130, 1 245, 1 375, 1 550, 1 680 nm) using in the TN detector developed by our team to verify the soil particle size correction method proposed in this paper. The results showed that the corrected 2.0, 0.9, 0.45 mm and original soil absorbance were reduced by 62%, 74%, 111% and 61%, respectively. It showed that the soil particle size correction method could reduce the interference of soil particle size to discrete NIR spectral data. Finally, BPNN was used to establish the TN models with different absorbance data. The results showed that the R2_v of the corrected soil absorbance model was improved by 25% compared with the original absorbance model. In summary, the soil particle size correction method proposed in this paper reduce the interference of the soil particle size on the discrete NIR spectral data, and improve the detection accuracy of the vehicle-mounted TN detector.