集成建模变量优选和参数学习的SVR盐渍化监测

SVR Salinization Monitoring based on Integrated Feature Subset Selection and Model Parameter Learning

当前基于机器学习算法反演土壤盐分含量(Soil Salt Content,SSC)较少关注模型参数和建模变量的优选。基于Sentinel-1 SAR、Landsat 8 OLI、DEM数据提取8类共40个环境变量,经Pearson相关分析初步筛选出候选特征变量,分别带入格网搜索(Grid Search,GS)算法、遗传算法(Genetic Algorithm,GA)、粒子群算法(Particle Swarm Optimization,PSO)同步优选支持向量回归(Support Vector Regression,SVR)的模型参数和建模变量,然后建立盐渍化监测模型(GSSVR、GA-SVR、PSO-SVR),选择最优模型反演玛纳斯灌区生长季SSC分布。结果表明:提取的环境变量与SSC相关性较好,植被指数和特征空间对盐渍化更为敏感;与GS-SVR相比,GASVR和PSO-SVR减少了建模变量,提高了模型精度,适应度值分别提高了53.87%、69.96%;生长季内,春秋季积盐,夏季脱盐,SSC均值变化趋势:整个研究区、中部和南部为降低—增加;北部为增加—降低—增加;依据生长季SSC小提琴图表明整个研究区,中部和北部SSC数值区间变化趋势为扩张—收缩—扩张,南部为扩张—收缩—平稳。

Salinization is one of the main forms of land degradation which leads to fragile ecological environment and low efficiency of agricultural production.Remote sensing combined with machine learning algorithm is one the most popular methods in salinization monitoring.In terms of machine learning algorithm,the model feature subset and parameters is vital to modeling accuracy.Therefore,accurate identification and optimization of model parameters and feature subset is crucial for machine learning based inversion and prediction of Soil Salt Content(SSC).Based on Sentinel-1 SAR,Landsat 8 OLI images and DEM data,a total of 40 environmental factors of 8 categories were extracted.In conjunction with Pearson correlation analysis,the Candidate Feature Variables(CFVs)were initially selected.The CFVs were introduced into the Grid Search(GS)algorithm,Genetic Algorithm(GA)and the Particle Swarm Optimization(PSO)to simultaneously identify the model parameter and feature subset of Support Vector Regression(SVR).Salinization monitoring models(GS-SVR,GA-SVR,PSO-SVR)were established,respectively.The optimal model was applied into the SSC prediction of Manasi Irrigation District in growing season,2016.The results show that the extracted environmental factors showed good correlations with SSC,and the vegetation indices and feature spaces were more sensitive to salinization than other types of environmental factors.Compared with GS-SVR,the GA-SVR and PSO-SVR methods improved the accuracy of the salinization monitoring while reducing the number of feature subset,and the fitness value increased by 53.87%and 69.96%,respectively.During the growing season,salt accumulates in spring and autumn and fades in summer.The trend of average SSC of the whole study area and the central part and the southern part was decreasing-increasing,while the northern part was increasing-decreasing-increasing.According to the SSC violin plots in the growing season,it was found that the trend of SSC range of the whole study area and the central part and the northern part was expansion-contraction-expansion,while it was expansion-contraction-stability in southern part of study area.This study provided the technical support for accurate salinization monitoring and dynamic change of SSC in growing season.