考虑质地影响的热带地区土壤有机质可见光—近红外反演方法协同优化研究

Co-optimization of a texture-conscious visible-near-infrared inversion model for tropical soil organic matter

关键词:  土壤有机质  可见光-近红外光谱  质地划分  光谱预处理  特征选择  随机森林  偏最小二乘回归

Key words:  soil organic matter  visible-near-infrared spectroscopy  texture classification  spectral pre-processing  feature selection  random forest  partial least squares regression

摘要:

【目的】 在充分考虑土壤质地差异的前提下，系统分析光谱预处理、特征选择与建模方法在 VIS-NIR 土壤有机质（SOM）反演中的协同优化效应，揭示其对模型精度、稳定性及跨质地泛化能力的影响规律。【方法】 文章将270个海南省典型农用地土壤样本按质地划分为砂土、壤土和黏土3类；对获取的土壤反射光谱，分别采用标准正态变量变换（SNV）和标准正态变量变换加自适应基线矫正（SNVBA）两种方法进行预处理；使用竞争性自适应重加权采样（CARS）进行特征波段筛选；分别构建了偏最小二乘回归（PLSR）与随机森林（RF）模型，采用10折交叉验证评估模型精度。【结果】 （1）Vis-NIR能有效预测SOM含量，但在PLSR模型性能受质地类型显著影响，总体上，壤土与黏土的预测精度明显高于砂土，说明质地划分能够提升线性模型适配性与稳定性。（2）复合预处理方法（SNVBA）显著改善了壤土与黏土的光谱表现，增强了模型的低反射特征解析力。（3）CARS对线性模型PLSR具有一定优化作用，但对非线性模型RF贡献较小。在所有质地与预处理组合中，RF模型表现最优且最稳定，其整体R2均保持在较高水平（>0.85），显著优于PLSR模型。【结论】 该研究提出的“质地划分+ SNVBA预处理+ RF建模”协同优化路径，能够有效削弱质地差异带来的干扰，提升光谱-有机质关系的稳定性，实现SOM的高精度反演。该研究建立的“质地划分—预处理适配—建模方法优选”一体化建模框架，为热带典型农田多质地条件下的土壤有机质快速监测与区域碳循环评估提供了可推广的技术支撑。

Abstract:

[Purpose] This study aims to systematically analyze the synergistic optimization effects of spectral preprocessing，feature selection，and modeling methods on visible-near infrared（Vis-NIR）soil organic matter（SOM）inversion while fully considering soil texture differences，and to reveal their impacts on model accuracy，stability，and cross-texture generalization ability.[Method] A total of 270 soil samples collected from typical agricultural lands in Hainan Province were classified into three texture types：sandy soil，loam，and clay. The acquired soil reflectance spectra were preprocessed using standard normal variate transformation（SNV）and a combined method of standard normal variate transformation with adaptive baseline correction（SNVBA）. Competitive adaptive reweighted sampling（CARS）was applied for characteristic wavelength selection. Partial least squares regression（PLSR）and random forest（RF）models were constructed，and model performance was evaluated using ten-fold cross-validation.[Result] （1）Vis-NIR spectroscopy effectively predicted SOM content；however，the performance of PLSR models was significantly influenced by soil texture. Overall，prediction accuracy for loam and clay was markedly higher than that for sandy soil，indicating that texture stratification improved the adaptability and stability of linear models. （2）The composite preprocessing method（SNVBA）significantly enhanced the spectral representation of loam and clay soils and improved the resolution of low-reflectance spectral features. （3）CARS contributed to the optimization of the linear PLSR model but provided limited improvement for the nonlinear RF model. Among all texture and preprocessing combinations，the RF model showed the best and most stable performance，with overall R2 values consistently exceeding 0.85 and significantly outperforming the PLSR model.[Conclusion] The proposed synergistic optimization strategy combining texture stratification，SNVBA preprocessing，and RF modeling effectively reduces the interference caused by soil texture heterogeneity，enhances the stability of the spectral-SOM relationship，and enables high-accuracy SOM inversion. This study establishes an integrated modeling framework of "texture stratification-preprocessing adaptation-model selection"，providing a transferable technical basis for rapid SOM monitoring and regional carbon cycle assessment in tropical agricultural soils with diverse texture conditions.