结合分数阶微分技术与机器学习算法的土壤有机碳含量光谱估测

Combination of Fractional Order Differential and Machine Learning Algorithm for Spectral Estimation of Soil Organic Carbon Content

采集新疆渭干河-库车河典型绿洲96个表层土样,测量其光谱反射率和土壤有机碳(SOC)含量,采用分数阶微分技术(阶数的取值范围为0~2,步长为0.2)结合极限学习机、随机森林、多元自适应回归样条函数、弹性网络回归和梯度提升回归树(GBRT)5种机器学习算法,并对SOC含量进行高精度估算。实验结果表明:分数阶微分的预处理效果优于整数阶微分;特定波段处相关性得到明显提高,最大相关性提高了0.220;作为集成学习的GBRT(验证集中决定系数为0.878,相对分析误差为3.142)在不同阶数下均优于其他模型,建议使用基于1.6阶光谱反射率的GBRT估测干旱区绿洲SOC含量。总之,基于可见光-近红外(VIS^NIR)结合分数阶微分技术与机器学习算法,为提高估测干旱区绿洲SOC含量的模型精度提出新方案。

In this study,96 surface soil samples are obtained from the typical oasis of the Ugan-Kuqa River in the Xinjiang Uyghur Autonomous Region and their spectral reflectance and soil organic carbon(SOC)content are evaluated.Using fractional order differential technique(with an order value range of 0-2 and a step size of 0.2)is combined with five machine learning algorithms,including the extreme learning machine,random forest,multiple adaptive regression spline function,elastic network regression,and gradient lifting regression tree(GBRT)algorithms,and high-precision estimation of SOC content.The experimental results show that the pretreatment effect obtained using a fractional order differential is better than that obtained using an integer order differential.The correlation at a specific band is significantly improved,and the maximum correlation is enhanced by approximately 0.220,In case of the GBRT,the verification concentration determination coefficient is 0.878 and the relative analysis error is 3.142,indicating that this type of integrated learning is superior to other models of different orders.GBRT based on a 1.6-order spectral reflectance should be used to estimate the SOC content of the oasis in arid areas.Thus,a new scheme based on the combination of visible light-near infrared(VIS-NIR)with the fractional order differential technology and machine learning algorithms is proposed in this study to improve the accuracy of thc model used for estimating the SOC content of the oasis in arid areas.