基于卫星和无人机遥感数据尺度转换的土壤盐渍化监测研究

Research on Soil Salinization Monitoring Based on Scale Conversion of Satellite and UAV Remote Sensing Data

为提高卫星遥感对土壤盐渍化的监测精度,以内蒙古河套灌区沙壕渠灌域内5块地为研究区,利用GF-1卫星遥感和无人机多光谱遥感分别获取2018年6月中旬的遥感影像数据,同步采集0~20 cm,20~40 cm深度的土壤样点,并引用洛伦兹曲线的原理以表征土壤异质性,同时引入BP神经网络(Back Propagation,BP)、支持向量机(Support Vector Machine,SVM)和极限学习机(Extreme Learning Machine,ELM)构建土壤盐渍化监测模型。采用重采样尺度转换方法,对无人机数据进行尺度上推,用尺度上推后的无人机数据修正GF-1卫星数据,对修正后的数据进行反演建模并与直接采用卫星数据建立的模型进行对比。结果表明:实验区异质性大小与变异系数大小呈正相关。无人机数据构建的机器学习算法模型精度高于卫星数据。其中20 cm深度下无人机遥感数据反演土壤含盐量的最优模型为SVM模型,决定系数(R^2)为0.875,均方根误差(RMSE)为0.132,相对分析误差(RPD)为2.773;40 cm深度下无人机遥感数据反演土壤含盐量的最优模型为BP模型R^2为0.709,RMSE为0.144,RPD为1.781;20 cm深度下GF-1卫星遥感数据反演土壤含盐量的最优模型为SVM模型,R^2为0.453,RMSE为0.245,RPD为0.055;40 cm深度下GF-1卫星遥感数据反演土壤含盐量的最优模型为BP模型R^2为0.271,RMSE为0.267,RPD为0.001。通过升尺度转换,可提高卫星遥感反演土壤盐分的模型精度,R^2可提高0.4~0.5,RMSE可减小0.061,RPD可提高1.308。可为改进卫星遥感监测土壤盐渍化方法提供参考。

In order to improve the monitoring accuracy of satellite remote sensing on soil salinization,GF-1 satellite remote sensing and Unmanned Aerial Vehicle multi-spectral remote sensing were used to obtain remote sensing image data in mid-June 2018 and simultaneously collect 0~20,20~40 cm in-depth soil salinity data.Through analysis,the principle of Lorentz curve was used to characterize soil heterogeneity.The BP neural network,support vector machine and extreme learning machine were introduced to construct soil salinization monitoring model.The resampling scale conversion method was used to scale up the UAV data,and the GF-1 satellite data was corrected with the scaled up UAV data.Then inversion modeling was carried out and compared with the model established by directly using satellite data.The results showed that the heterogeneity of the experimental area was positively correlated with the coefficient of variation.The accuracy of the machine learning algorithm model constructed by drone data was higher than that of satellite data.Among them,the optimal model for inversion of soil salt content from UAV remote sensing data at a depth of 20 cm was the SVM model,R^2 was 0.875,RMSE was 0.132,and RPD was 2.773;the optimal model for inverting soil salinity from UAV remote sensing data at a depth of 40 cm was BP model,R^2 was 0.709,RMSE was 0.144,and RPD was 1.781;the optimal model for retrieving soil salt content from GF-1 satellite remote sensing data at a depth of 20 cm was the SVM model,R^2 was 0.453,RMSE was 0.245,and RPD was 0.055;the optimal model for inversion of soil salinity from GF-1 satellite remote sensing data at a depth of 40 cm was BP model,R^2 was 0.271,RMSE was 0.267,and RPD was 0.001.Through upscaling,the accuracy of the model of soil salinity inversion from satellite remote sensing can be improved.The R^2 can be increased by 0.4 to 0.5,the RMSE can be reduced by 0.061,and the RPD can be increased by 1.308.This study can provide a reference for improving the method of monitoring soil salinization by satellite remote sensing.