Abstract: The alkali⁃hydrolyzable nitrogen（AH⁃N）content in soil profiles reflects the short⁃term nitrogen supply potential of soil. However，its hyperspectral estimation is often hindered by high⁃dimensional redundant features，which limits model accuracy. To optimize the combination of methods and improve estimation accuracy，this study focused on the distribution area of fluvo⁃aquic soil in Henan Province.Specifically，11 soil profiles（1 m depth）were selected across farmlands，orchards and vegetable fields.A near⁃infrared hyperspectral imager was utilized to acquire hyperspectral images of the profiles，obtaining the spectral information of 220 soil samples. Five spectral preprocessing methods，such as standard normal variate（SNV）and first derivative（FD），were used in combination with competitive adaptive reweighted sampling（CARS），successive projections algorithm（SPA），and uninformative variable elimination（UVE）to extract wavelengths sensitive to AH⁃N. Partial least squares regression（PLSR）and least squares support vector machine（LS⁃SVM）models were constructed to compare and analyze the estimation accuracy of different method combinations.The results indicated that both SNV and FD preprocessing improved the model performance. Overall，wavelength selection methods yielded better performance when coupled with the LS⁃SVM model than with the PLSR model. Among them，the FD⁃CARS⁃LS⁃SVM model exhibited the optimal performance，with a prediction set R² of 0.89 ，a root mean square error（RMSE）of 11.58 mg/kg，and a relative prediction deviation（RPD）of 2.92.Profile validation based on the optimal model showed that the R² values for all profiles ranged from 0.90 to 0.98，with RMSE values between 4.33 and 11.77 mg/kg，and RPD values all exceeding 2.3.These results demonstrate the robustness and stability of the model and indicate its capability for accurate inversion and vertical distribution characterization of AH⁃N content in fluvo⁃aquic soil profiles. In conclusion，the integration of FD preprocessing and CARS variable selection effectively eliminates redundant information in hyperspectral data. When coupled with the nonlinear LS⁃SVM model，this approach achieves optimal estimation of AH⁃N content in fluvo⁃aquic soil profiles. Furthermore，the proposed method effectively characterizes the vertical spatial differentiation of AH⁃N.

Key words: Fluvo?aquic soil, Alkali?hydrolyzable nitrogen, Characteristic band, Model, Hyperspectrum

摘要： 土壤剖面碱解氮含量反映土壤短期供氮潜力，但其高光谱估测容易受到数据高维冗余特征制约，影响模型精度。为优化方法组合提升估测精度，以河南省潮土分布范围为研究区，选取农田、果园和蔬菜地共11个1 m深土壤剖面，利用近红外高光谱成像仪采集剖面高光谱图像，获取220个土壤样品的光谱信息。采用标准正态变换（SNV）、一阶微分（FD）等5种光谱预处理方法，结合竞争性自适应重加权采样（CARS）、连续投影法（SPA）和无信息变量消除（UVE）提取碱解氮敏感波段，分别构建偏最小二乘回归（PLSR）模型与最小二乘支持向量机（LS-SVM）模型，对比分析不同方法组合的估测精度。结果表明，SNV和FD预处理均提升了模型预测性能，特征波段筛选方法在LS-SVM模型中的表现总体优于PLSR模型。其中，FD-CARS-LS-SVM模型表现最优，其预测集R²达0.89，均方根误差（RMSE）为11.58 mg/kg，相对分析误差（RPD）为2.92。基于最优模型进行剖面验证发现，所有剖面的R2介于0.90~0.98，RMSE介于4.33~11.77 mg/kg，RPD均大于2.3，表明该模型稳定性高，可实现潮土剖面碱解氮含量的精准反演与纵向分布表征。综上，FD预处理结合CARS变量筛选可有效消除高光谱冗余信息，结合非线性LS-SVM模型可实现潮土剖面碱解氮含量最优估测，该方法可有效表征碱解氮在剖面纵向上的空间分异规律。

关键词: 潮土, 碱解氮, 特征波段, 模型, 高光谱