摘要
以黑龙江农田黑土为研究对象,利用遗传算法(GA)波长选择结合偏最小二乘法(PLS)回归建立土壤有机碳(SOC)的预测模型。通过设定以下GA参数:波长选择数量上限k、初始种群大小P及迭代次数N,采用单点优化方式逐一确定各参数。结果表明,在主成份数为7的情况下,当GA的参数取N=300、P=300、k=50时,GA模型最优;模型的校正决定系数R2=0.922、校正均方根误差RMSEC=1.74、交叉检验均方根误差RMSECV=1.80;模型的预测决定系数R2=0.931、预测均方根误差RMSEP=1.84、预测相对误差RPD=3.81。与原始光谱的PLS模型相比,R2由0.900提升至0.922,RPD由3.38提升至3.81。结果表明,通过GA进行波长选择能够优化模型,提升模型稳定性以及预测精确性。
Partial least squares(PLS) regression combined with genetic algorithm(GA), which was a general variable selection technique, was used to establish the prediction model of soil organic carbon(SOC) in this paper. We set the following GA parameters: maximum Number of wavelength selection(k), population size(P), and the Number of iterations(N). These parameters were determined using single point optimization. The results showed that, under the principal component 7, when the parameters of GA were as follows: N=300, P=300, k=50, the GA model was optimal,and the coefficient of determination of calibration R2 was 0.922, root mean square error of calibration(RMSEC) was1.74, root mean square error of cross validation(RMSECV) was 1.80, coefficient of determination of prediction R2was0.931, root mean square error of prediction(RMSEP) was 1.84, and the residual prediction deviation(RPD) was 3.81.Compared with the original PLS regression model, the coefficient of determination R2 was improved from 0.900 to0.922, while RPD increased from 3.38 to 3.81. It suggested that GA used in the wavelength selection enhanced the robustness and accuracy of the prediction model.
出处
《土壤通报》
CAS
CSCD
北大核心
2014年第4期795-800,共6页
Chinese Journal of Soil Science
基金
国家自然科学基金(41171199)
中国科学院战略性先导科技专项"应对气候变化的碳收支认证及相关问题(XDA05050501)"资助
关键词
遗传算法
近红外光谱
土壤有机碳
偏最小二乘法
Genetic algorithm
Near infrared spectroscopy
Soil organic carbon
Partial least squares
