结合分数阶微分和异常值识别的土壤有机质高光谱反演研究

Hyperspectral inversion of soil organic matter combining with fractional differentiation and outlier recognition

关键词:  土壤有机质  蒙特卡洛交叉验证  分数阶微分  偏最小二乘回归

Key words:  soil organic matter  Monte Carlo cross validation  fractional differentiation  partial least squares regression

摘要:

【目的】 结合分数阶微分和异常值识别，提高土壤有机质模型反演精度，实现土壤有机质含量的快速、准确估计。【方法】 文章以吉林省伊通县黑土区为研究区，基于实地采集的213个土壤样本和HyMap-C机载高光谱传感器获取高光谱影像，选择S-G函数和分数阶微分进行光谱预处理，竞争性自适应重加权采样（Competitive Adaptive Reweighted Sampling，CARS）提取特征波段建立土壤有机质含量偏最小二乘回归（Partial Least Squares Regression，PLSR）反演模型，并使用蒙特卡洛交叉验证（Monte Carlo Cross-Validation，MCCV）进行异常值识别。【结果】 （1）将分数阶微分用于机载高光谱可以放大光谱特征，阶数越高、特征越明显，低阶分数微分对噪音不敏感；（2）CARS方法能有效压缩光谱信息；全样本建模中0.4阶分数阶微分CARS-PLSR建模表现较优，但总体精度仍然不高；（3）使用MCCV剔除异常值后，0.6阶分数阶微分CARS-PLSR建立的土壤有机质含量反演模型精度最高，训练集和测试集的均方误差分别为0.219 %和0.207 %，决定系数分别为0.745和0.823，相对分析误差为2.246；（4）将模型应用于高光谱影像中，可获得研究区高精度、高分辨率土壤有机质空间分布图。【结论】 分数阶微分可以实现光谱增强，具有数据挖掘潜力，同时识别并剔除异常值可以有效优化高光谱反演模型，为使用高光谱影像进行土壤有机质含量反演工作提供新思路。

Abstract:

【Purpose】 Combining fractional differentiation and outlier recognition to improve the accuracy of the soil organic matter inversion model and achieve fast and accurate estimation of the soil organic matter.【Method】 This research taked the black soil region in Yitong County，Jilin Province as the research area. Based on 213 soil samples collected in the field and hyperspectral imagery obtained from the HyMap-C airborne hyperspectral sensor，the S-G function and fractional-order differentiation were used for spectral preprocessing，the Competitive Adaptive Reweighted Sampling（CARS）algorithm was employed to extract characteristic bands and establish a Partial Least Squares Regression（PLSR）model for soil organic matter estimation. An Monte Carlo Cross-Validation（MCCV）method was used to identify outliers.【Result】 （1）Using fractional differentiation for airborne hyperspectral could amplify spectral features. High fractional order will make the features more prominent，while low order fractional differentiation was not sensitive to noise；（2）The CARS method could effectively compress spectral information；In the full sample modeling，the 0.4-order fractional differentiation CARS-PLSR model demonstrated superior performance，but the overall accuracy was still limited；（3）After removing outliers using MCCV，the 0.6 FOD-CARS-PLSR soil organic matter inversion model exhibited the highest accuracy，with root mean square errors（RMSE）of 0.219% and 0.207% for the training and testing sets，respectively. The coefficient determination（R2）values were 0.745 and 0.823，and the relative percent deviation （RPD）was 2.246；（4）Applying the model to hyperspectral images can obtain high-precision and high-resolution spatial distribution maps of soil organic matter in the study area.【Conclusion】 Fractional differentiation can enhance spectral features and holds potential for data mining. Simultaneously，the identification and removal of outliers effectively optimize the hyperspectral inversion model，offering novel insights for the use of hyperspectral imagery in soil organic matter content inversion.