【作者】 张玉华；

【导师】 杨磊； 杨东凯；

【作者基本信息】 山东农业大学 ， 农业工程与信息技术（专业学位）， 2021， 硕士

【摘要】 土壤湿度是用来描述土壤干湿程度的一个物理量,反应了农作物的水分供应状况,准确监测土壤湿度是实现农业稳产、高产的重要基础。对农业、气象和全球水循环等领域的应用意义重大具有重要意义。全球导航卫星系统反射信号(Global Navigation Satellite System-Reflectometry,GNSS-R)技术作为一种低成本的遥感技术,近年来受到广泛关注。本文利用全球导航卫星系统直射与反射信号两者间的干涉现象以及获取土壤湿度的理论,通过随机森林(Random Forest,RF)和深度神经网络(Deep Neural Networks,DNN)两种算法,构建了基于随机森林的GNSS-IR土壤湿度反演模型和基于深度神经网络的GNSS-IR土壤湿度反演模型,并将模型数据处理结果分别与线性回归模型处理结果、实测数据结果进行对比分析;最后通过地基实验进行了验证。本文以GPS PRN(Pseudo Random Noise)9卫星为例对实验结果进行展示,并将均方根误差(Root Mean Square Error,RMSE)作为卫星反演结果评价指标进行对比分析,得出最终结论,主要研究内容和结果如下:(1)构建基于随机森林的GNSS-IR土壤湿度反演模型利用随机森林算法对全球定位系统(Global Positioning System,GPS)L1、L2两个频段的信噪比(Signal to Noise Ratio,SNR)频率、幅度、相位等观测量进行单频单变量和单频双变量处理并建立反演模型,将结果与线性回归处理结果进行对比分析。以PRN 9为例,L1频段频率、幅度、相位单变量的RMSE分别降低了62.07%、49.09%、67.02%;L2频段频率、幅度、相位单变量的RMSE分别降低了68.55%、71.22%、64.07%,结果表明随机森林算法在算法复杂度较小的情况下,在单变量土壤湿度反演方面,取得了良好的效果,但在双频数据融合方面表现有所欠缺。(2)构建基于深度神经网络的GNSS-IR土壤湿度反演模型利用深度神经网络分别对GPS L1、L2两个频段的SNR频率、幅度、相位等观测量分别进行单频单变量和单频双变量土壤湿度反演并建立模型。与随机森林算法结果相比,以PRN 9为例,处理L1频段频率、幅度、相位单变量的RMSE分别降低了69.70%、67.86%、50.79%;L2频段的RMSE分别降低了48.72%、30.00%、53.33%。处理L1频段双变量的RMSE分别降低了85.37%、95.59%、95.89%;L2频段的RMSE分别降低了96.49%、87.88%、98.73%。结果表明深度神经网络算法取得了比随机森林更好的效果,特别是多变量数据融合方面相比随机森林更加有效,但其算法复杂度相对于随机森林算法较高,且对双变量进行土壤湿度反演更有效。本文通过随机森林算法和深度神经网络算法分别建立反演模型,与线性回归方法处理结果和实验实测数据进行比较。结果表明机器学习可以有效的提高土壤湿度反演精度,推动了GNSS-R在农业生产方面的广泛应用。更多还原

【Abstract】 Soil moisture is a physical quantity used to describe the degree of soil dryness and wetness,which reflects the water supply of crops.Accurate monitoring of soil moisture is an important basis for realizing stable and high agricultural yield.It is of great significance to the application of agrometeorology and global water cycle.Global Navigation Satellite SystemReflectometry(GNSS-R)technology,as a low-cost remote sensing technology,has attracted wide attention in recent years.This paper is based on the theoretical study of soil moisture acquisition by using the interference between direct and reflected signals of the global navigation satellite system.Then two algorithms,random forest and deep neural networks,were used to construct GNSS-IR soil moisture inversion model respectively,and compared and analyzed the data processing results of linear regression model and measured data.Finally,ground-based experiments were carried out to verify the results,and GPS PRN 9 was taken as an example to demonstrate the experimental results,and the RMSE was used as the evaluation index of satellite inversion results,and the final conclusion was obtained.The main research contents and results of this paper are as follows.(1)Established a GNSS-IR soil moisture inversion model based on random forestRandom forest algorithm was used to process the SNR frequency,amplitude and phase of two frequency bands of the GPS L1 and L2.Then single parameter and dual-parameters processing were carried out,and the inversion model was established,and the results were compared with linear regression processing.Taking PRN 9 as an example,the RMSE of frequency,amplitude and phase single parameter in L1 decreases 62.07%,49.09%,67.02%respectively,and the RMSE of frequency,amplitude and phase single parameter in L2 decreases 68.55%,71.22%,64.07%,respectively.The results show that the random forest algorithm has achieved good results in single parameter soil moisture inversion under the low algorithm complexity,but the performance in dual-parameter data fusion is insufficient.(1)Established a GNSS-IR soil moisture inversion model based on deep neural networkDeep neural network was used to process the SNR frequency,amplitude and phase of two frequency bands of the GPS L1 and L2.Then single parameter and dual-parameters processing were carried out,and the inversion model was established,and the results were compared with random forest algorithm processing.Taking PRN 9 as an example,the RMSE of frequency,amplitude and phase single parameter in L1 decreases 69.70%,67.86%,50.79%respectively,and the RMSE of frequency,amplitude and phase single parameter in L2 decreases 48.72%,30.00%,53.33% respectively.And the RMSE of frequency,amplitude and phase dual-parameter in L1 decreases 85.37%,95.59%,95.89% respectively,and the RMSE of frequency,amplitude and phase dual-parameter in L2 decreases 96.49%,87.88%,98.73%respectively.The results show that the deep neural network algorithm achieves better results than the random forest,especially the dual-parameter data fusion is more effective than the random forest algorithm,but its algorithm complexity is higher than the random forest algorithm,and the dual-parameter soil moisture inversion is more effective.In this paper,the random forest algorithm and deep neural network algorithm were used to establish the inversion model respectively,and the results of linear regression method and experimental measured data were compared.The results show that machine learning can effectively improve the accuracy of soil moisture inversion,and promote the wide application of GNSS-R in agricultural production.更多还原