精细农业变量施肥取决于对农田的土壤养分分布的了解, 快速获取土壤信息是实施精细农业的基础。 速效钾是土壤肥力的重要参数, 是植物生长发育所必需的营养元素。 对土壤速效钾含量进行测量, 是了解土壤肥力的重要途径, 是实现精细农业的必要条件。 以山西典型褐土土壤为研究对象, 采集农田耕层褐土土壤样品共169份, 样品经风干处理, 手动捏碎较大的土粒并去除杂质后, 未经研磨过筛处理而直接用于土壤近红外高光谱的测量。 根据实验室速效钾含量测定结果, 将所有土壤样品分为两类: 其中速效钾含量低于100 mg·kg^-1的样品共144个, 随机选取108个作为低含量建模集( L_c), 剩余36个作为低含量验证集( L_p); 速效钾含量高于100 mg·kg^-1的样品共25个, 随机选取19个作为高含量建模集( H_c), 剩余6个作为高含量验证集( H_p)。 其中 L_c和 H_c统称为所有含量建模集( T_c), L_p和H_p统称为所有含量验证集( T_p)。 获取所有土壤样本9501 650 nm范围内的近红外高光谱图像。 分别采用平均光谱曲线(R)、 平均光谱曲线的一阶导数(FD)、 平均光谱曲线与一阶导数共同建模(R&FD)、 平均光谱曲线与一阶导数的乘积(R*FD)、 平均光谱曲线与一阶导数的商(R/FD)等五种光谱数据预处理方法, 结合偏最小二乘法(PLS), 分别对建模集 T_c, L_c及 H_c建模, 然后分别对验证集 T_p, L_p及 H_p进行验证。 结果表明: 土壤的平均光谱反射率随速效钾含量的增大呈现先增加后减小的趋势。 当速效钾含量低于100 mg·kg^-1时, 所有波段的光谱反射率随速效钾含量的增加而增加; 当速效钾含量在100200 mg·kg^-1之间时, 所有波段的光谱反射率均达到最大值。 当速效钾含量超过200 mg·kg^-1时, 9501 400 nm的光谱反射率急剧减小, 但曲线的整体斜率显著增加; 且速效钾含量越高, 曲线整体斜率越大。 当速效钾含量高于100 mg·kg^-1时, 平均光谱曲线的一阶导数显著增大, 且随速效钾含量的增加而增加。 该研究建立的PLS模型, 可以对整体(所有速效钾含量)和高含量(≥100 mg·kg^-1)速效钾进行有效预测, 但无法对低含量(≤100 mg·kg^-1)速效钾进行预测。 建模效果最好的光谱预处理方法为R*FD, 其次为FD, R, 而R&FD, R/FD预测效果相对较差。 最优建模方式为: R*FD结合 T_c建模, 其PLS主因子个数为2个, RMSE_c=29.293, RPD_c=4.669,Rc2=0.956; 对 T_p的验证效果为RMSE_p=29.438, RPD_p=4.740,Rp2=0.958; 对 H_p的验证效果为RMSE_p=23.033, RPD_p=3.199,Rp2=0.915。 该模型能够根据土壤速效钾的含量对土壤进行分类: 当预测值小于100 mg·kg^-1时, 表明土壤速效钾含量低于100 mg·kg^-1, 具体含量不确定; 当预测值大于100 mg·kg^-1时, 预测值则能够很好反映土壤速效钾的真实含量。 由于选用的土壤样本未经研磨和过筛处理, 因而能够大大缩短样本制备时间, 提高预测效率。 该研究结果可为近红外高光谱成像应用于褐土土壤除速效钾含量以外其他营养成份的快速预测提供参考。

Variable rate fertilization in precision agriculture depends on the understanding of the distribution of soil nutrients in the farmland. The rapid acquisition of soil information is the basis for the application of precision agriculture. Available potassium is an important parameter of soil fertility, and it is a necessary nutrient element for plant growth. The measurement of the content of available potassium in soil is an important way to understand the soil fertility, and it is a precondition of the realization of precision agriculture. In this paper, a total of 169 farmland plough cinnamon soil samples were collected in Shanxi province. All samples were air dried, with the larger soil particles crumbled and impurities removed manually and directly used for measuring soil near infrared hyperspectral without grinding and sieving. According to the measuring results of the available potassium content in the laboratory, all soil samples were divided into two parts. There were 144 samples of available potassium content less than 100 mg·kg^-1, and 108 samples were randomly selected as the low content modeling sets ( L_c), and the remaining 36 samples as the low content validation sets ( L_p). There were 25 samples of available potassium content more than 100 mg·kg^-1, and 19 samples were randomly selected as the high content modeling sets ( H_c), and the remaining 6 samples as the high content validation sets ( H_p). L_c and H_c were collectively known as all modeling sets ( T_c), and L_p and H_p as all validation sets ( T_p). Near infrared hyperspectral imaging technology was used to obtain near infrared hyperspectral images in the range of 9501 650 nm of all soil samples. There were five different spectral data preprocessing methods used in this paper: the average spectral curve (R), the first derivative of the average spectral curve (FD), the average spectral curve and the first derivative co-modeled (R&FD), the product of the average spectral curve and the first derivative (R*FD) and the quotient of average spectral curve and first derivative (R/FD). Combined with partial least squares (PLS) method, the models were built using the modeling set T_c, L_c and H_c respectively. The validation sets T_p, L_p and H_p were verified respectively. The results showed that: along with the increase of available potassium content, the average spectral reflectance of soil increased first and then decreased. When the content of available potassium was less than 100 mg·kg^-1, the spectral reflectance of all bands increased with the increase of available potassium content. When the content of available potassium was between 100200 mg·kg^-1, the spectral reflectance of all bands reached the maximum. When the available potassium content was more than 200 mg·kg^-1, the spectral reflectance of 9501 400 nm decreased sharply, but the overall slope of the curve increased significantly. The higher the available potassium content was, the larger the overall slope of the curve was. When the content of available potassium was higher than 100 mg·kg^-1, the first derivative of the average spectral curve increased significantly, and increased with the increase of available potassium content. The PLS models proposed in this paper could predict the whole (all available potassium content) and high content (≥100 mg·kg^-1) of available potassium effectively; but could not predict the low content (≤100 mg·kg^-1) of available potassium. The best spectral preprocessing method was: R*FD, followed by FD and R. The predict results of R&FD and R/FD were relatively poor. The optimal modeling method was R*FD combined with T_c. The number of PLS principal factors was 2, RMSE_c=29.293, RPD_c=4.669,Rc2=0.956; RMSE_p=29.438, RPD_p=4.740,Rp2=0.958 for the validation sets of T_p; RMSE_p=23.033, RPD_p=3.199,Rp2=0.915 for the validation sets of H_p. This model could classify soil according to the content of available potassium. When the predicted value was less than 100 mg·kg^-1, it indicated that the content of available potassium in soil was less than 100 mg·kg^-1, and the specific content was uncertain; while when the predicted value was higher than 100 mg·kg^-1, the predicted value could reflect the real content of soil available potassium well. Because the soil samples selected in this paper were used without ground or sifted, the time of sample preparation could be greatly shortened and the prediction efficiency could be improved greatly. The results in this study can provide a reference for the rapid prediction of nutrients including available potassium content in cinnamon soil using near infrared hyperspectral imaging technology.