基于Boruta算法和GA优化混合地统计模型的土壤有机质空间分布预测

Predicting the spatial distribution of soil organic matter using the model consisting of the Boruta algorithm and the optimized GA combined with the geostatistical method

建立土壤有机质(SOM)空间预测模型不仅可以准确预测SOM含量的空间分布,而且对科学化土壤管理和完善生态系统服务具有重要意义。本文以山西省临汾市永和县土壤为研究对象,从数字高程模型(DEM)和植被遥感数据中提取出地形因子和植被指数,结合土壤本身属性为变量因子,采取Boruta算法从变量因子中筛选出与SOM相关性强的特征变量为辅助变量作为模型输入,实测SOM值作为模型输出,选择普通克里格方法(OK)、反向传播神经网络(BPNN)、遗传算法优化的BP神经网络(GA-BPNN)和GA优化BP神经网络结合地统计方法(GA-BPNN-OK)对训练集样本SOM含量进行预测,并利用验证集样本对比分析预测精度。研究结果显示:Boruta算法优选出特征变量并且对其进行了重要性排列,依次为:全氮>地形湿度指数(TWI)>高程>坡度>归一化植被指数(NDVI)>增强型植被指数(EVI);4种方法对SOM的预测结果虽然局部会有差异,但整体的空间分布基本一致,在研究区内呈现出西部和西南部地区低、东部和东南部地区高的空间分布趋势,与其他3种模型相比,GA-BPNN-OK模型预测的SOM分布图对低值区和高值区的划分更加明显、细致;预测精度指标对比得出,GA-BPNN-OK法的均方根误差(RMSE为0.059)、平均绝对误差(MAE为0.240)、平均相对误差(MRE为0.165)最小,且拟合系数(R2为0.78)最高。同时为了验证采用Boruta算法对模型精度有所提高,将全变量与特征筛选之后的变量作为GA-BPNN法的模型输入,对预测结果进行对比,结果表明采取Boruta算法后模型误差减小。因此采取Boruta算法筛选出特征变量作为辅助变量,GA-BPNN-OK法对于SOM含量空间分布的精度最高,两者结合为最优预测模型。

Establishing a spatial prediction model for soil organic matter(SOM)can accurately predict the spatial distribution of SOM content,playing a significant role in scientific soil management and ecosystem service enhancement.Focusing on the soils in Yonghe County,Linfen City,Shanxi Province,this study extracted topographic factors and vegetation indices from the digital elevation model(DEM)and vegetation remote sensing data.With soil attributes as variable factors,this study,using the Boruta algorithm,selected the characteristic variablescorrelating strongly with SOM from variable factors as auxiliary variables.These auxiliary variables were used as model inputand the measured SOM values as model output.The SOM content in samples in the training set was predicted usingthe ordinary Kriging(OK)method,the back propagation neural network(BPNN),the genetic algorithm-optimized BPNN(GA-BPNN),and the improved BPNN combined with the geostatistical method(the GA-BPNN-OK method)separately.The prediction accuracy was comparatively analyzed based on samples in the validation set.The results show that:(1)The Boruta algorithm ranked the selected characteristic variables in order of importance,obtaining the sequence of total nitrogen>topographic wetness index(TWI)>elevation>slope>normalized difference vegetation index(NDVI)>enhanced vegetation index(EVI);(2)Despite local differences,the SOM prediction results obtained using the four methods exhibited roughly the same overall spatial distribution:low in the western and southwestern portions of the study areabut high in the eastern and southeastern portions;(3)Compared to the other three models,the GA-BPNN-OK model demonstrated more distinct low-and high-value areas in the predicted SOM distribution.(4)As revealed by the comparison of prediction accuracy indices,the GA-BPNN-OK method yielded a minimum root mean square error(RMSE)of 0.059,a minimum mean absolute error(MAE)of 0.240,a minimum mean relative error(MRE)of 0.165,and a maximum fitting coefficient(R 2)of 0.78.To verify the effects of the Boruta algorithm in improving model accuracy,global variables,as well as the variables determined through characteristic selection,were used as the model inputof the GA-BPNN method.The comparison of the prediction results indicates that the Boruta algorithm reduced the model error.Therefore,the Boruta algorithm and the GA-BPNN-OK method constitute the optimal prediction model for the spatial distribution of SOM content.