中国矿业大学环境与测绘学院何佳星:粒子群优化卷积神经网络GNSS-IR土壤湿度反演办法 |《测绘学报》2023年52卷第8期
<div style="color: black; text-align: left; margin-bottom: 10px;">
<div style="color: black; text-align: left; margin-bottom: 10px;">
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/6d06975bdc3d9aba1ebf1a1e0a93faab~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725647938&x-signature=gL77vxdSDo5Pw0%2BzENvAwnCfjjs%3D" style="width: 50%; margin-bottom: 20px;"></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">本文内容<span style="color: black;">源自</span>于《测绘学</span><span style="color: black;">报》2023年第8期(</span><span style="color: black;">审图号GS京(2023)1524号</span><span style="color: black;">)</span></span></p><strong style="color: blue;"><span style="color: black;">粒子群优化卷积神经网络GNSS-IR土壤湿度反演<span style="color: black;">办法</span></span></strong>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;"><span style="color: black;">何佳星1</span></strong><strong style="color: blue;"><span style="color: black;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/793a0ccd4be03810f6431ac75de5a701~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725647938&x-signature=yElml49onWArVMZKzMybpEXaKbg%3D" style="width: 50%; margin-bottom: 20px;"></span></strong><strong style="color: blue;"><span style="color: black;">, 郑南山1,2<img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/594d6323e16971ca8e8d3c5f77514622~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725647938&x-signature=v92L0%2FUXcXgERxPUdozSn8xr2Qo%3D" style="width: 50%; margin-bottom: 20px;"></span></strong><strong style="color: blue;"><span style="color: black;">, 丁锐1,2, 张克非1,2, 陈天悦1 </span></strong></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">1. 中国矿业大学环境与测绘学院, 江苏 徐州 221116;</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">2. 中国矿业大学自然资源部国土环境与灾害监测重点实验室, 江苏 徐州 221116</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;"><span style="color: black;">基金项目:</span></strong><span style="color: black;">国家自然科学基金(41974039);国家自然科学基金联合重点(U22A20569);自然资源部国土环境与灾害监测重点实验室开放基金(LEDM2021B11);国家重点<span style="color: black;">开发</span>计划课题(2019YFC1805003);江苏省<span style="color: black;">科研</span>生<span style="color: black;">研究</span>与实践创新计划(KYCX22_2594);中国矿业大学<span style="color: black;">科研</span>生创新计划(2022WLJCRCZL253)</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;">摘要</strong><span style="color: black;">:</span><span style="color: black;"><span style="color: black;">全世界</span>卫星导航系统干涉<span style="color: black;">测绘</span>法(GNSS-interferometric reflectometry, GNSS-IR)是一种新兴对地观测遥感技术, 利用该<span style="color: black;">办法</span>能实现土壤湿度监测, <span style="color: black;">拥有</span>很高的应用<span style="color: black;">潜能</span>。</span><span style="color: black;">针对土壤湿度反演的建模问题, 本文构建一种集成粒子群优化算法(particle swarm optimization, PSO)和卷积神经网络(convolutional neural network, CNN)的GNSS-IR土壤湿度反演模型, 将多颗GPS卫星两个频点信噪比(signal-to-noise ratio, SNR)观测数据提取的特征参数<span style="color: black;">做为</span>模型输入, <span style="color: black;">经过</span>粒子群算法求解卷积神经网络超参数, 对模型进行优化实现高精度反演。</span><span style="color: black;">以P041站点为例<span style="color: black;">仔细</span>描述了模型<span style="color: black;">创立</span>过程, 本文<span style="color: black;">办法</span>的均方根误差为0.015 0, 相较于基于单星线性、多星线性、未优化CNN和BP神经网络模型分别降低约60%、27%、31%和21%;并<span style="color: black;">经过</span><span style="color: black;">位置于</span><span style="color: black;">区别</span>地理环境的COPR、P183、P341站点验证模型的<span style="color: black;">靠谱</span>性和适用性。</span><span style="color: black;"><span style="color: black;">实验</span>结果<span style="color: black;">显示</span>, 融合多源观测数据<span style="color: black;">创立</span>PSO优化CNN的GNSS-IR土壤湿度反演模型, 能有效反演土壤湿度, <span style="color: black;">必定</span>程度上<span style="color: black;">控制</span>了<span style="color: black;">区别</span>下垫面环境的影响, <span style="color: black;">拥有</span>较强的适用性。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">关键词:土壤湿度 GNSS-IR SNR 卷积神经网络 粒子群</span> </p><span style="color: black;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/40cf8d7c07dda1ec51a5c20525b957a3~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725647938&x-signature=FIWGhn1yIH%2BKE1oIpbEDc0taqqM%3D" style="width: 50%; margin-bottom: 20px;"></span><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/d1ec5c1967f306110d683e6cafd2ccb4~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725647938&x-signature=0pmMIho6VFrsR1hoPn8Nm1%2BdZig%3D" style="width: 50%; margin-bottom: 20px;">
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">何佳星, 郑南山, 丁锐, 等. 粒子群优化卷积神经网络GNSS-IR土壤湿度反演<span style="color: black;">办法</span>. 测绘学报,2023,52(8):1286-1297. DOI: </p>10.11947/j.AGCS.2023.20220277
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">HE Jiaxing, ZHENG Nanshan, DING Rui, et al. A GNSS-IR soil moisture inversion method based on the convolutional neural network optimized by particle swarm optimization. Acta Geodaetica et Cartographica Sinica, 2023, 52(8): 1286-1297. DOI: 10.11947/j.AGCS.2023.20220277</span></p><strong style="color: blue;"><span style="color: black;">阅读全文</span></strong><span style="color: black;">:http://xb.chinasmp.com/article/2023/1001-1595/20230806.htm</span>
<h2 style="color: black; text-align: left; margin-bottom: 10px;"><strong style="color: blue;"><span style="color: black;">引 言</span></strong><span style="color: black;"><span style="color: black;">土壤湿度是气候系统中重要的生态水文变量,对<span style="color: black;">科研</span>农业生产、气候变化和<span style="color: black;">全世界</span>水循环等有着<span style="color: black;">非常</span>重要的科学<span style="color: black;">道理</span>[<span style="color: black;">1</span>]。恒温<span style="color: black;">叫作</span>重法、时域反射仪和电阻法等传统土壤湿度<span style="color: black;">测绘</span><span style="color: black;">办法</span>,<span style="color: black;">测绘</span>精度较高,但<span style="color: black;">测绘</span>范围极小,不满足环境监测、预防灾害的<span style="color: black;">需要</span>;可见光和热红外遥感<span style="color: black;">办法</span>能<span style="color: black;">长时间</span>监测土壤水分,但易受到天气、上层植被等<span style="color: black;">原因</span>影响;被动微波遥感<span style="color: black;">办法</span>对土壤介电常数<span style="color: black;">敏锐</span>,但难以实现小范围区域的高精度监测;SAR反演土壤湿度<span style="color: black;">办法</span>有较高时空分辨率,但后向散射系数与土壤水分之间的关系还有待<span style="color: black;">科研</span>[<span style="color: black;">2</span>-<span style="color: black;">3</span>]。新兴的GNSS-R(GNSS-Reflectometry)技术最早于20世纪90年代提出[<span style="color: black;">4</span>-<span style="color: black;">5</span>],利用GNSS系统<span style="color: black;">供给</span>的<span style="color: black;">全世界</span>覆盖、高时空分辨率卫星信号,<span style="color: black;">经过</span>分析处理反射/散射信号极化特性、频率、相位、振幅等信息能有效反映地表<span style="color: black;">理学</span>参数变化,是当前土壤湿度监测<span style="color: black;">办法</span>的重要<span style="color: black;">弥补</span>。<span style="color: black;">日前</span>该技术<span style="color: black;">导致</span>大地<span style="color: black;">测绘</span>学界关注,并被广泛用于对地观测遥感[<span style="color: black;">6</span>-<span style="color: black;">10</span>],在水位观测、雪深探测、土壤湿度监测等方面取得一系列成果,其中刘经南院士指出GNSS-R<span style="color: black;">科研</span>的关键是<span style="color: black;">加强</span>反演<span style="color: black;">制品</span>的精度[<span style="color: black;">11</span>]。</span></span><span style="color: black;"><span style="color: black;">关于GNSS-R遥感,2002年美国航空航天局首次开展了机载GPS单极化天线接收反射信号遥感土壤和植被参数<span style="color: black;">实验</span>(soil moisture experiment 2002, SMEX02),2008年文献首次提出利用GNSS-IR技术反演土壤湿度,并<span style="color: black;">显示</span>多径振幅与土壤湿度之间存在<span style="color: black;">必定</span>的<span style="color: black;">关联</span>性。在传统的单星线性和单星指数反演模型中,<span style="color: black;">广泛</span>假定<span style="color: black;">实验</span>场地为理想状态,但反射信号与<span style="color: black;">目的</span><span style="color: black;">周边</span>环境密切<span style="color: black;">关联</span>,<span style="color: black;">目的</span>反射、散射特征会随环境<span style="color: black;">原因</span>改变,<span style="color: black;">引起</span>特征参数难以准确获取,<span style="color: black;">作为</span>GNSS-IR土壤湿度反演<span style="color: black;">科研</span>的热点问题[<span style="color: black;">13</span>-<span style="color: black;">14</span>]。文献<span style="color: black;">思虑</span>反射面有效高度随时间变化,将低高度角(2°~30°)与高高度角(30°~70°)结合,<span style="color: black;">经过</span>归一化和反转<span style="color: black;">得到</span>新的信噪比时间序列,从而<span style="color: black;">加强</span>反演精度。文献利用南非萨瑟兰站2008—2014年数据,<span style="color: black;">创立</span>土壤湿度的长时序动态监测,并指出L2C波段反演精度略高于L1和L2P波段反演精度,信号采样率由30 s<span style="color: black;">提高</span>到1 s对土壤湿度探测的准确度影响很小。文献提出一种半经验信噪比模型,在无先验知识<span style="color: black;">状况</span>下,直接从信噪比序列中重构直反射信号,能有效<span style="color: black;">控制</span>噪声从而<span style="color: black;">加强</span>反演精度。文献采用熵值法将L1和L2两个波段数据融合,与实测土壤湿度<span style="color: black;">创立</span>经验模型,大幅<span style="color: black;">加强</span>了反演精度。在多星融合土壤湿度反演<span style="color: black;">科研</span>方面,文献提出一种监测土壤湿度的多星线性回归模型,<span style="color: black;">处理</span>单颗卫星极易<span style="color: black;">显现</span>跳变的问题,<span style="color: black;">加强</span>突发性降雨时段的土壤湿度反演精度。文献提出基于多星融合的LS-SVM土壤湿度探测<span style="color: black;">办法</span>,认为土壤湿度反演是个非线性过程,发挥<span style="color: black;">设备</span>学习的<span style="color: black;">优良</span>,有效整合各卫星信息,有利于土壤湿度精确探测。在探究植被和地表对土壤湿度反演影响的<span style="color: black;">科研</span>中,文献利用植被状态变化与信噪比指标变化之间的关系,<span style="color: black;">创立</span>了顾及植被影响的土壤湿度反演模型,有效<span style="color: black;">控制</span>植被对土壤湿度探测的影响。文献 <span style="color: black;">创立</span>了粗糙度修正的土壤湿度反演模型,并指出当粗糙度均方根在0.010~0.025 m时,修正模型反演精度较高。在<span style="color: black;">设备</span>学习辅助反演的<span style="color: black;">科研</span>中,文献利用BP神经网络<span style="color: black;">创立</span>归一化振幅与NDVI之间的关系,有效改进线性回归模型不足,<span style="color: black;">提高</span>反演精度。文献构建一个<span style="color: black;">包括</span>5个隐含层每层200个神经元的神经网络模型,在沿海地区取得了良好的风速反演性能。文献<span style="color: black;">一样</span>利用人工神经网络模型反演土壤湿度,将卫星观测值和地表辅助参数<span style="color: black;">做为</span>输入,生成高分辨率的土壤湿度预测网。</span></span><span style="color: black;"><span style="color: black;"><span style="color: black;">针对</span>单天线GNSS接收机而言,接收的反射信号总会受大气、地表等环境<span style="color: black;">原因</span>影响,<span style="color: black;">怎样</span><span style="color: black;">控制</span>和削弱干扰<span style="color: black;">原因</span>影响是土壤湿度反演的关键。然而多种地表环境<span style="color: black;">原因</span>互相耦合,难以定量描述<span style="color: black;">每一个</span><span style="color: black;">原因</span>对GNSS-R技术反演精度的影响。使得<span style="color: black;">怎样</span>利用丰富的多频多星观测信息,并<span style="color: black;">思虑</span>非线性<span style="color: black;">原因</span>的<span style="color: black;">功效</span>,<span style="color: black;">作为</span>目前土壤湿度反演的难点。<span style="color: black;">因此呢</span>,本文<span style="color: black;">经过</span><span style="color: black;">实验</span>分析<span style="color: black;">选择</span>土壤湿度反演的<span style="color: black;">靠谱</span>特征参数,发挥深度学习驱动模型的<span style="color: black;">优良</span>,<span style="color: black;">创立</span>多径干涉特征参数和土壤湿度间的隐式特征关系模型,以减轻下垫面特征变化等<span style="color: black;">原因</span>的影响,并采用粒子群优化算法<span style="color: black;">加强</span>深度学习模型的泛化能力,<span style="color: black;">提高</span>GNSS-IR土壤湿度反演模型的精度和适应性。<span style="color: black;">选取</span>美国板块边界观测计划(plate boundary observatory, PBO)<span style="color: black;">供给</span>的GPS卫星L1、L2波段观测数据,将相邻气象站<span style="color: black;">供给</span>的实测土壤湿度数据<span style="color: black;">做为</span>参考,在缺乏先验信息(地表粗糙度、植被覆盖、植被含水量等)<span style="color: black;">要求</span>下,利用PSO寻找最优解速度快、CNN自主学习能力强的优点,以隐含式<span style="color: black;">办法</span>构建多径干涉特征参数与土壤湿度间的<span style="color: black;">繁杂</span>映射关系。<span style="color: black;">经过</span>一个站点展开<span style="color: black;">仔细</span>建模,3个站点模型验证,并与多种模型对比分析来<span style="color: black;">评估</span>本文<span style="color: black;">办法</span>的反演精度和适应性。</span></span></h2>
<h2 style="color: black; text-align: left; margin-bottom: 10px;"><strong style="color: blue;"><span style="color: black;">1 GNSS陆面参数反演<span style="color: black;">基本</span></span></strong></h2>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">1.1 数据<span style="color: black;">源自</span></span></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">PBO计划是美国科学项目地球透镜计划的<span style="color: black;">构成</span>部分,<span style="color: black;">重点</span>用于<span style="color: black;">科研</span>北美大陆西部板块边缘等地区地壳形变,共设置了1000多台大地<span style="color: black;">测绘</span>型高精度GNSS接收机。本文<span style="color: black;">选择</span>PBO<span style="color: black;">供给</span>的P041<span style="color: black;">实验</span>场地(39.949 5°N,105.194 3°W)、COPR<span style="color: black;">实验</span>场地(34.414 9°N,119.879 5°W)、P183<span style="color: black;">实验</span>场地(38.313 7°N,123.068 9°W)和P341<span style="color: black;">实验</span>场地(40.650 7°N,122.606 9°W)为<span style="color: black;">科研</span>对象,<span style="color: black;">思虑</span><span style="color: black;">实验</span><span style="color: black;">科研</span>所需各类数据的完备性,<span style="color: black;">选取</span>观测时段均为2013年DOY1—DOY365。P041站点周边地势开阔平坦,植被类型为天然杂草,植被影响较小,且无大型遮挡物,如图 1所示,有利于实现GNSS-IR土壤湿度反演[<span style="color: black;">26</span>],其余3个站点周边环境较<span style="color: black;">繁杂</span>,如COPR和P341站点周边存在植被影响,P183站点周边存在地形影响。4个站点架设的Trimble接收机由钢制三脚架固定支撑,并为用户<span style="color: black;">供给</span>15 s采样间隔的GPS观测数据(</span><span style="color: black;">https://www.unavco.org</span><span style="color: black;">)。站点旁均有气象站<span style="color: black;">供给</span>0.05 m深度的实地土壤湿度数据。以P041站点<span style="color: black;">仔细</span>介绍本文<span style="color: black;">办法</span>的参数<span style="color: black;">选择</span>,建模过程和<span style="color: black;">实验</span>结果,其余3个站点用于验证本文<span style="color: black;">办法</span>。</span></p>
<div style="color: black; text-align: left; margin-bottom: 10px;">
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/56430977ea20db23910c5351083d7fc8~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725647938&x-signature=sLI%2FQUMWIm9xQmDkebJXaw9eAxw%3D" style="width: 50%; margin-bottom: 20px;"><span style="color: black;"><span style="color: black;">图 1 </span>4个站点的周边环境<span style="color: black;">Fig. 1</span> Surrounding environment of 4 sites</span>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">图选项</span></p>
</div>
</div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">1.2 GNSS-IR遥感反演陆面参数原理</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">因为</span>多路径效应的产生,接收机天线会<span style="color: black;">同期</span>接收直射信号和反射信号并在两者间产生<span style="color: black;">必定</span>的相位差,几何关系如图 2(a)所示,经叠加干涉后形成图 2(b)所示的复合信号。图中θ<span style="color: black;">暗示</span>卫星高度角;h<span style="color: black;">暗示</span>仪器高度;SNR<span style="color: black;">暗示</span>复合信号Ac的信噪比值;Ad和Ar分别<span style="color: black;">暗示</span>直射信号和反射信号;ψ<span style="color: black;">暗示</span>直反信号的相位差。</span></p>
<div style="color: black; text-align: left; margin-bottom: 10px;">
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/f12497a3a0368f570004b513fbb4f6a2~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725647938&x-signature=jkzWd59L71%2FNZ3WBjoVc081tJ2Q%3D" style="width: 50%; margin-bottom: 20px;"><span style="color: black;"><span style="color: black;">图 2 </span>地基GNSS-IR原理<span style="color: black;">Fig. 2</span>Schematic diagram of ground-based GNSS-IR</span>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">图选项</span></p>
</div>
</div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">如图 3(a)所示,在低卫星高度角下,卫星信号受多路径影响<span style="color: black;">明显</span>,SNR震荡幅度大,且呈周期性变化。采用低阶多项式拟合直射信号,从SNR序列中分离并提取出反射信号,如图 3(b)所示,文献指出剔除直射信号后的反射分量与卫星高度角正弦间存在<span style="color: black;">必定</span>周期的正弦或余弦函数关系,表达式为</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/66dde8e8629a2c305ef8e55100d75f19~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725647938&x-signature=lEbkwWIWmZBzZFX3gi8NN4Pe%2BoA%3D" style="width: 50%; margin-bottom: 20px;"><span style="color: black;"> <span style="color: black;">(1)</span></span></p>
<div style="color: black; text-align: left; margin-bottom: 10px;">
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p26-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/be670bc1efb6055397a1afbf43a3b69a~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725647938&x-signature=s%2FnMf6adFBo4B4MmahKSDyUbsis%3D" style="width: 50%; margin-bottom: 20px;"><span style="color: black;"><span style="color: black;">图 3 </span>信噪比预处理分析<span style="color: black;">Fig. 3</span> SNR preprocessing analysis</span>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">图选项</span></p>
</div>
</div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">式中,DSNR<span style="color: black;">暗示</span>信噪比残差序列;A<span style="color: black;">暗示</span>反射信号振幅;H<span style="color: black;">暗示</span>天线相位中心距反射面的垂直距离,采用Lomb-Scargle频谱分析得到;λ<span style="color: black;">暗示</span>载波波长;φ<span style="color: black;">暗示</span>相位偏移。按式(1)给出的余弦函数关系,对图 3(b)中的信噪比残差曲线进行非线性最小二乘拟合,可求出反射分量的特征参数A和φ。<span style="color: black;">经过</span><span style="color: black;">创立</span>振幅、相位、反射面高度与土壤湿度之间的关系,实现土壤湿度的反演。</span></p>
<h2 style="color: black; text-align: left; margin-bottom: 10px;"><strong style="color: blue;"><span style="color: black;">2 PSO优化CNN土壤湿度反演算法</span></strong></h2>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">2.1 PSO-CNN集成算法</span></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">卷积神经网络<span style="color: black;">重点</span>由输入层、卷积层、下采样层(池化层)、全连接层和输出层<span style="color: black;">形成</span>,其基本结构如图 4所示。实质上,CNN是一种前馈式神经网络,采用前向传播和反向传播算法进行训练和网络优化,<span style="color: black;">最后</span>求得输出值。前向传播是自下而上的过程,输入层的数据经卷积层的卷积和池化层处理,将提出的特征向量传入全连接层,当输出结果与期望相符时,输出结果。网络输出的结果与期望值不符时,进行反向传播,<span style="color: black;">按照</span>损失计算梯度,更新每一层权值,是自上而下的过程[<span style="color: black;">28</span>]。</span></p>
<div style="color: black; text-align: left; margin-bottom: 10px;">
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/abb7f67f8d8458f98a52827b4a65349e~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725647938&x-signature=w6kZwhhdsHLkFIV8hG9omrhYJD4%3D" style="width: 50%; margin-bottom: 20px;"><span style="color: black;"><span style="color: black;">图 4 </span>PSO优化CNN网络结构过程<span style="color: black;">Fig. 4</span> Improved convolutional neural network structure with particle swarm optimization</span>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">图选项</span></p>
</div>
</div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">CNN网络训练前需设置<span style="color: black;">关联</span>参数,即超参数设置,<span style="color: black;">重点</span><span style="color: black;">包含</span>卷积核的数量和<span style="color: black;">体积</span>、激活函数类型、学习率和梯度下降算法等。超参数的设定直接决定了神经网络算法的准确性和收敛速度,<span style="color: black;">因此呢</span>需要一种算法寻找CNN的最优超参数。粒子群优化算法是文献受人工生命<span style="color: black;">科研</span>结构的启发,<span style="color: black;">经过</span>模拟鸟觅食过程提出一种迭代的群优化算法。其优点在于简单易行、收敛速度快、参数设置少,粒子群优化算法求解CNN最佳网络结构过程如图 4所示。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">粒子群优化卷积神经网络算法基本过程如下。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">(1) 设定最大迭代次数、种群规模及其他基本参数。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">(2) 初始化或更新粒子位置和速度。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">(3) 将粒子<span style="color: black;">做为</span>CNN的结构,随机初始化CNN权重,进行网络训练。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">(4) 以模型预测值和参考值的RMSE作为适应度,计算公式(2)中适应度Fitness,并确定局部最优和历史最优</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/a5891766ecc95dc0843214010e7ea9e3~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725647938&x-signature=uO6IEO6cX5Bthb%2FiWHpaqXP8hz4%3D" style="width: 50%; margin-bottom: 20px;"><span style="color: black;"> <span style="color: black;">(2)</span></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">式中,yi为模型预测值;</span><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/fb42199ec93ac320505585390c75e600~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725647938&x-signature=lZM4CEUHdx4SKzciPcw18cnAIAU%3D" style="width: 50%; margin-bottom: 20px;"><span style="color: black;">为参考值。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">(5) <span style="color: black;">小于</span>设定最小RMSE<span style="color: black;">或</span>达到最大迭代次数时结束;否则返回<span style="color: black;">过程</span>(2)。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">2.2 土壤湿度优化模型构建<span style="color: black;">办法</span></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">按1.2节中<span style="color: black;">办法</span>,从去除趋势项的信噪比残差曲线中提取得到特征参数,文献指出特征参数与土壤湿度呈现<span style="color: black;">必定</span>的<span style="color: black;">关联</span>性,但<span style="color: black;">因为</span>站点<span style="color: black;">实质</span><span style="color: black;">状况</span><span style="color: black;">区别</span>,<span style="color: black;">引起</span>特征参数与土壤湿度之间<span style="color: black;">关联</span>性存在差别,应<span style="color: black;">按照</span>测区地表环境、卫星信号质量等信息,<span style="color: black;">选择</span>恰当的多源数据集<span style="color: black;">做为</span>网络的输入值。<span style="color: black;">因此呢</span>,在两者存在线性关系的<span style="color: black;">基本</span>上,把特征参数<span style="color: black;">做为</span>CNN预测模型的输入在理论上是可行的。输入的数据集为</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/3b9b9fcd59291e4e138d8324ce51bd06~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725647938&x-signature=R8K%2Fm%2Fvisj5z0Gqs6RnWb7Tw9%2B8%3D" style="width: 50%; margin-bottom: 20px;"><span style="color: black;"> <span style="color: black;">(3)</span></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">式中,X<span style="color: black;">暗示</span>输入CNN的<span style="color: black;">全部</span>数据集;Xn<span style="color: black;">暗示</span>单个特征参数数据集;n<span style="color: black;">暗示</span>有效特征参数数量,为正整数。其中Xn<span style="color: black;">暗示</span>为</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/e8363144ce8f477e181d2b559418a6d4~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725647938&x-signature=SLqDlktpXEnUGwIyNXW%2FCK1dlLE%3D" style="width: 50%; margin-bottom: 20px;"><span style="color: black;"> <span style="color: black;">(4)</span></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">式中,m<span style="color: black;">暗示</span>各颗卫星年积日长度,为正整数。为了消除指标的量纲影响,使数据梯度变化<span style="color: black;">显著</span>,加快梯度下降求最优解的速度,对特征参数值和土壤湿度参考值进行归一化处理</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/fad260869f9a37e4232683b130dfe7c2~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725647938&x-signature=vIfUUOk5fJp90nRHzB06yNcDI7Y%3D" style="width: 50%; margin-bottom: 20px;"><span style="color: black;"> <span style="color: black;">(5)</span></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">式中,y<span style="color: black;">暗示</span>归一化后的值;x<span style="color: black;">暗示</span>归一化前的值;max<span style="color: black;">暗示</span>向量最大值;min<span style="color: black;">暗示</span>向量最小值。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">CNN两个卷积层中卷积核<span style="color: black;">体积</span>设为1×1,激活函数类型均为relu,梯度下降<span style="color: black;">办法</span><span style="color: black;">选取</span>sgdm,其余CNN超参数及其初始值范围见表 1所示,超参数最优解由粒子群算法求解得到,每颗粒子在迭代过程中,<span style="color: black;">经过</span>跟踪两个极值来更新自己,在找到这两个最优值时,粒子<span style="color: black;">按照</span>式(6)更新自己的速度和位置</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/daa03a862dc0289650452b75ac3efa08~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725647938&x-signature=CtDfQnD5Z64Xy%2F0DbzVTaFxVKkE%3D" style="width: 50%; margin-bottom: 20px;"><span style="color: black;"> <span style="color: black;">(6)</span></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">表 1</span> CNN超参数初始值范围设置<span style="color: black;">Tab. 1</span> CNN hyperparameter initial value range setting</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/e81169161bf8d396cc225005978cb208~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725647938&x-signature=fp7DNppHw37qABJJRRii4uJT04E%3D" style="width: 50%; margin-bottom: 20px;"></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">式中,vidk<span style="color: black;">暗示</span>第k次迭代中粒子i速度的第d维分量;xidk<span style="color: black;">暗示</span>第k次迭代中粒子i位置的第d维分量;pbestidk<span style="color: black;">暗示</span>粒子i局部最优值的第d维分量;gbestdk<span style="color: black;">暗示</span>全局最优值的第d维分量;ω<span style="color: black;">暗示</span>惯性权重;c1和c2<span style="color: black;">暗示</span>学习因子;r1和r2是0~1的随机数。假设优化<span style="color: black;">目的</span>为了找到最小值,则pbestidk和gbestdk的更新方式为</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/3f03dd963c9e76fc1f9558561c9ea975~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725647938&x-signature=N4DLq2WVj4A2pTTO5rtvWGyBmLM%3D" style="width: 50%; margin-bottom: 20px;"><span style="color: black;"> <span style="color: black;">(7)</span></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/de23b2c822fc91604bedd32d5ad3bdca~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725647938&x-signature=440ms2XPpL3hJpWRdLmWmBeOM5Q%3D" style="width: 50%; margin-bottom: 20px;"><span style="color: black;"> <span style="color: black;">(8)</span></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">式中,f(·)是<span style="color: black;">目的</span>函数;N为粒子群中粒子的总数。为避免数据偶然性,更好验证模型精度,本文采用K折交叉验证开展<span style="color: black;">实验</span>,将原始数据集分成K份,每次从K份中取出<span style="color: black;">区别</span>的一部分<span style="color: black;">做为</span>测试集,其余K-1份<span style="color: black;">做为</span>训练集,<span style="color: black;">保证</span>每份数据均被用作测试集,最后把K次<span style="color: black;">实验</span>结果取平均,<span style="color: black;">做为</span><span style="color: black;">实验</span><span style="color: black;">最后</span>结果。图 5为K折交叉验证的基本过程。</span></p>
<div style="color: black; text-align: left; margin-bottom: 10px;">
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/6bceeb69865cb3bff6c24f89c8d4eb82~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725647938&x-signature=GCLqhiZ0NnD5Z6cNzlwM%2FTWf35g%3D" style="width: 50%; margin-bottom: 20px;"><span style="color: black;"><span style="color: black;">图 5 </span>K折交叉验证<span style="color: black;">Fig. 5</span>K-fold cross-validation</span>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">图选项</span></p>
</div>
</div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">综上可得,粒子群优化卷积神经网络的土壤湿度反演算法可分为3个<span style="color: black;">重点</span><span style="color: black;">过程</span>,基本过程如图 6所示。</span></p>
<div style="color: black; text-align: left; margin-bottom: 10px;">
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/69d7f809143eca5c4cf88fe0f37cac42~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725647938&x-signature=tCFlY4ozE3sDA%2BI6cLFtS6l3fr0%3D" style="width: 50%; margin-bottom: 20px;"><span style="color: black;"><span style="color: black;">图 6 </span>PSO优化CNN土壤湿度反演算法<span style="color: black;">Fig. 6</span> PSO-CNN soil moisture inversion algorithm</span>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">图选项</span></p>
</div>
</div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">(1) 多径干涉特征参数提取。从原始SNR数据中分离出反射SNR分量,截取低高度角数据,利用LSP频谱分析求得特征参数h,采用非线性最小二乘<span style="color: black;">办法</span>得到特征参数A、φ。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">(2) 参数归一化和模型初始化。设置PSO算法基本参数和CNN超参数初始值范围,以归一化后的多颗卫星数据集<span style="color: black;">做为</span>CNN的输入。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">(3) 模型网络训练和结果输出。以模型预测值和土壤湿度参考值的均方根误差为<span style="color: black;">目的</span>函数,当误差<span style="color: black;">少于</span>阈值或达到最大迭代次数后网络训练结束,输出<span style="color: black;">最后</span>结果。</span></p>
<h2 style="color: black; text-align: left; margin-bottom: 10px;"><strong style="color: blue;"><span style="color: black;">3 优化算法反演模型构建与分析</span></strong></h2>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">3.1 站点数据预处理分析</span></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">GNSS-IR技术利用低高度角卫星数据实现地表参数的反演,文献给出了5°~30°的<span style="color: black;">举荐</span>高度角范围,但其与<span style="color: black;">实质</span>地表<span style="color: black;">状况</span>和卫星数据质量<span style="color: black;">相关</span>,由图 3(b)可知, P041站点高度角5°~20°内信噪比残差曲线震荡<span style="color: black;">拥有</span>周期性且更加稳定,部分卫星5°~30°、5°~25°的信噪比数据<span style="color: black;">包括</span>了<span style="color: black;">更加多</span>噪声,对其进行LSP频谱分析时易产生双峰问题,分析结果如图 7所示,<span style="color: black;">因此呢</span>设置5°~20°为P041站点的截止高度角。</span></p>
<div style="color: black; text-align: left; margin-bottom: 10px;">
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/8ad4e6baf085f5d6555516c912ade987~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725647938&x-signature=L8PtCCeHEkim5541RSFWB5eMjGo%3D" style="width: 50%; margin-bottom: 20px;"><span style="color: black;"><span style="color: black;">图 7 </span>PRN 11<span style="color: black;">区别</span>高度角范围信噪比序列的LSP<span style="color: black;">Fig. 7</span> LSP about SNR observations of PRN 11 for different elevation angles</span>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">图选项</span></p>
</div>
</div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">3.2 建模参数<span style="color: black;">选择</span>分析</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">经过</span>分析各特征参数对<span style="color: black;">目的</span>区域土壤湿度<span style="color: black;">敏锐</span>性以<span style="color: black;">选择</span>适合的建模参数。图 8为该站点降雨事件前后反射信号余弦曲线,表 2是降雨前后特征参数的变化<span style="color: black;">状况</span>,降雨后信噪比残差序列的余弦拟合曲线<span style="color: black;">显著</span>向右偏移,相位值随之减小,证明了相位值能响应土壤湿度的变化。而振幅和反射面高度在降雨前后<span style="color: black;">无</span>呈现<span style="color: black;">显著</span>规律,绝大<span style="color: black;">都数</span>的线性<span style="color: black;">关联</span>系数<span style="color: black;">少于</span>0.3,<span style="color: black;">显示</span><span style="color: black;">这次</span><span style="color: black;">实验</span>中两者对土壤湿度变化不<span style="color: black;">敏锐</span>。</span></p>
<div style="color: black; text-align: left; margin-bottom: 10px;">
<div style="color: black; text-align: left; margin-bottom: 10px;">
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/3587d9ca546f4d6f30a78873a123d8b8~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725647938&x-signature=s3oRe26PdvwVmPiZO62ABsbC%2FHM%3D" style="width: 50%; margin-bottom: 20px;"></p><span style="color: black;"><span style="color: black;">图 8 </span>降雨事件前后反射信号余弦函数曲线变化(以PRN01、PRN19、PRN27为例)<span style="color: black;">Fig. 8</span> Cosine function curve changes for PRN01, PRN19 and PRN27 before and after rainfall</span>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">图选项</span></p>
</div>
</div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">表 2</span> 降雨前后特征值变化<span style="color: black;">状况</span>(以PRN01、PRN19、PRN27为例)<span style="color: black;">Tab. 2</span>Changes of SNR metrics for PRN01, PRN19 and PRN27 before and after rainfall</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/6a59f54d61e079f8827c75eba32c642d~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725647938&x-signature=0ohDhSDesMcUdV0rLKAQvNDTrNc%3D" style="width: 50%; margin-bottom: 20px;"></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">另外</span>,为避免单星反射区域有限、有效时段较短、<span style="color: black;">反常</span>跳变等问题,本文将多颗卫星<span style="color: black;">区别</span>时段的L1、L2波段相位值<span style="color: black;">做为</span>反演模型的输入。经<span style="color: black;">实验</span>筛选后<span style="color: black;">选择</span>了36组相位值(来自PRN01、07、09、11、15、19、20、23、24和31共10颗卫星<span style="color: black;">区别</span>时段的L1、L2数据)。其中卫星<span style="color: black;">选择</span>遵循以下原则[<span style="color: black;">32</span>-<span style="color: black;">33</span>]。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">(1) LSP频谱分析质量。<span style="color: black;">因为</span>系统噪声、地表粗糙等<span style="color: black;">原由</span>,频谱分析时会产生多峰问题,<span style="color: black;">选取</span>多峰未<span style="color: black;">显现</span><span style="color: black;">或</span><span style="color: black;">显现</span>极少的卫星。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">(2) 卫星反射区域方位。<span style="color: black;">因为</span>每颗卫星轨道设计<span style="color: black;">区别</span>,卫星反射信号经过的区域会有差异,所选卫星的反射区域不宜太集中,应较为均匀地分布在站点四周。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">(3) 卫星相位反演效果。卫星反演精度受信号质量、地表环境等<span style="color: black;">原因</span>影响,<span style="color: black;">选取</span>相位与土壤湿度<span style="color: black;">关联</span>系数大于0.5的卫星。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">图 9(a)展示了10颗卫星<span style="color: black;">第1</span>菲涅尔反射区从5°至20°的整体变化<span style="color: black;">状况</span>。高度角为5°时,<span style="color: black;">第1</span>菲涅尔反射区均为长轴约50 m、短轴约4 m的椭圆,单颗卫星有效面积约为150 m2;高度角为20°时,椭圆长轴约为6 m、短轴约为2 m,单颗卫星有效反射面积约为9 m2。随着卫星高度角的增大,<span style="color: black;">第1</span>菲涅尔反射区域变得更小,更靠近天线。10颗卫星相位值与土壤湿度参考值回归分析结果如图 9(b)所示, <span style="color: black;">关联</span>系数集中在0.6~0.8之间,<span style="color: black;">关联</span>系数最高为0.842 2,最低为0.557 4,平均为0.721 7。</span></p>
<div style="color: black; text-align: left; margin-bottom: 10px;">
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/366ed5a2921793ef748276252c2cce9c~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725647938&x-signature=K4HXNk86J%2FE1angaTeO3ubgzmAI%3D" style="width: 50%; margin-bottom: 20px;"><span style="color: black;"><span style="color: black;">图 9 </span>P041站点10颗GPS卫星概况<span style="color: black;">Fig. 9</span> Overview of 10 GPS satellites at P041 site</span>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">图选项</span></p>
</div>
</div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">3.3 模型精度评定</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">为进一步克服地表环境差异性的影响、<span style="color: black;">加强</span>土壤湿度反演效果,利用PSO-CNN优化算法构建土壤湿度反演模型,以<span style="color: black;">以上</span>卫星多径干涉相位值<span style="color: black;">做为</span>CNN输入值,CNN超参数初始值范围见表 1所示。参照文献将PSO的最大迭代次数设置为40,种群规模为20,粒子维度为5,学习因子c1=c2=1.494 55,惯性权重ω=0.729,以模型预测值和土壤湿度参考值的均方根误差<span style="color: black;">做为</span>粒子群算法的适应度。<span style="color: black;">经过</span>局部最优和历史最优两个特殊值更新粒子位置,并<span style="color: black;">持续</span>向最优解位置<span style="color: black;">挨近</span>,<span style="color: black;">最后</span>解算得到最佳网络结构。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">将相位数据集划分成5份,进行5折交叉验证,每份数据都需<span style="color: black;">做为</span>一次网络训练的测试集,经过粒子群算法优化后的卷积神经网络取得最佳网络结构,表 3给出每次训练后网络得到的<span style="color: black;">详细</span>超参数。5次网络训练结果的回归分析如图 10所示,PSO-CNN模型的预测值与土壤湿度有良好的<span style="color: black;">关联</span>性,<span style="color: black;">关联</span>系数最高为0.966 7,最低为0.700 6,均方根误差<span style="color: black;">少于</span>0.022 8,能有效反映土壤湿度的变化<span style="color: black;">状况</span>。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">表 3</span> PSO优化网络后的CNN超参数值<span style="color: black;">Tab. 3</span> CNN hyperparameters after PSO optimized network</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><img src="https://p26-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/46f6d31e1828110ced07aeba8a3e13c2~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725647938&x-signature=sYvegT3nVeTq8C2mCrDWqAT6nXM%3D" style="width: 50%; margin-bottom: 20px;"></p>
<div style="color: black; text-align: left; margin-bottom: 10px;">
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/9c9b67ac2b2366204b57f4fcf6ac6bac~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725647938&x-signature=1vg1KwOmmYLzThkYe2O9Y%2BbJ6w4%3D" style="width: 50%; margin-bottom: 20px;"><span style="color: black;"><span style="color: black;">图 10 </span>5折交叉验证结果的回归分析<span style="color: black;">Fig. 10</span> Regression analysis of the 5-fold cross-validation</span>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">图选项</span></p>
</div>
</div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">为了对比分析PSO-CNN优化模型的性能效果,本文加入单星线性回归模型、多星线性回归模型、未优化的CNN模型和BP神经网络模型进行对比分析,其中未优化CNN模型的超参数采用常规<span style="color: black;">方法</span>,见表 3最后一列。4种模型<span style="color: black;">运用</span>与PSO-CNN模型相同组的数据进行5折交叉验证,其中多星线性回归模型、未优化CNN模型和BP神经网络模型输入与PSO-CNN模型一致。对比图 11中单颗卫星的反演结果(便于图表说明和对比分析,随机<span style="color: black;">选取</span>PRN 01、PRN 15和PRN 23共3颗卫星),结合图 12给出的结果分析可知,单星反演的效果有差异,其<span style="color: black;">原由</span>可能有两点:一是土壤湿度值取每日的平均值,每颗卫星经过站点上空的时间有差异,湿度参考值<span style="color: black;">不可</span>做到完全统一;二是卫星信号来自<span style="color: black;">区别</span>的方位角,<span style="color: black;">引起</span>反射信号经过的地表区域不完全一致。<span style="color: black;">因为</span><span style="color: black;">区别</span>卫星在时空上的差异性,仅采用单星反演<span style="color: black;">不可</span>精确监测土壤湿度。本文融合来自两个频点、各个方位、<span style="color: black;">区别</span>时间段的卫星反射信号,以多星的相位集<span style="color: black;">做为</span>CNN输入,并利用CNN<span style="color: black;">处理</span>相位与土壤湿度之间存在的非线性问题,实现土壤湿度高精度反演。</span></p>
<div style="color: black; text-align: left; margin-bottom: 10px;">
<div style="color: black; text-align: left; margin-bottom: 10px;">
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/3c7ff7a3bd21f4012136cf926e488970~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725647938&x-signature=8fgy5PxR0m9izt4mcA9Dap4kIVY%3D" style="width: 50%; margin-bottom: 20px;"></p><span style="color: black;"><span style="color: black;">图 11 </span>5种反演模型精度分析<span style="color: black;">Fig. 11</span> Accuracy analysis of 5 inversion models</span>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">图选项</span></p>
</div>
</div>
<div style="color: black; text-align: left; margin-bottom: 10px;">
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/356e54054ab13044b2ac3596c8e4b7ba~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725647938&x-signature=Hpxq3GK3vitfFqGZa2Xt68D0Dso%3D" style="width: 50%; margin-bottom: 20px;"><span style="color: black;"><span style="color: black;">图 12 </span>单星线性回归模型对比分析结果<span style="color: black;">Fig. 12</span> Comparison of single satellite regression models</span>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">图选项</span></p>
</div>
</div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">图 11左列中,PSO-CNN模型和其余4种模型均能对土壤湿度变化做出反应,但整体来看,PSO-CNN模型<span style="color: black;">针对</span>湿度变化更<span style="color: black;">敏锐</span>,与土壤湿度参考值一致性更好;由图 11右列各模型误差的箱形图<span style="color: black;">能够</span>看出,PSO-CNN模型的误差均值更靠近零,其箱体部分(整体排序25%~75%的部分)最集中,证明其误差波动最小,反演效果最佳。为了进一步<span style="color: black;">表现</span>PSO-CNN模型的反演精度,取5次交叉<span style="color: black;">实验</span>结果的平均值,对各模型进行误差分析和精度评定,以模型预测值和土壤湿度参考值之间的<span style="color: black;">关联</span>系数(R)、均方根误差(RMSE)、平均绝对误差(MAE)、平均相对误差(MRE)、最大上偏差(ES)和最大下偏差(EI)<span style="color: black;">做为</span>评定标准,结果分析见图 12和表 4所示。PSO-CNN模型预测值与土壤湿度参考值的<span style="color: black;">关联</span>系数达0.892 0,比单星线性回归模型<span style="color: black;">加强</span>约50%,比多星线性回归模型<span style="color: black;">加强</span>约6%,比未优化CNN模型<span style="color: black;">加强</span>约7%,比BP神经网络模型<span style="color: black;">加强</span>约4%;均方根误差方面,PSO-CNN模型为0.015,比单星线性回归模型低约60%,比多星线性回归模型低约27%,比未优化CNN模型低约31%,比BP神经网络模型低约21%;绝对误差方面,PSO-CNN模型为0.011 6,比单星线性回归模型降低约60%,比多星线性回归模型降低约37%,比未优化CNN模型降低约41%,比BP神经网络模型低约27%。单星线性模型精度最差,未优化的CNN模型精度与多星线性模型精度相当,BP神经网络模型精度稍好,PSO-CNN模型精度最高,<span style="color: black;">显示</span>卷积神经网络结合粒子群优化算法的有效性。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">表 4</span> 5种反演模型的性能指标对比<span style="color: black;">Tab. 4</span> Performance comparison of 5 inversion models</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/6cc5c3f4602b709d38ecd812350c4916~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725647938&x-signature=fwltV8uQyrMqqHqOZ2fNaESHovw%3D" style="width: 50%; margin-bottom: 20px;"></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">3.4 模型验证与适用性分析</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">为了进一步验证本文<span style="color: black;">办法</span>的<span style="color: black;">靠谱</span>性并探究其在<span style="color: black;">区别</span>下垫面环境下的适用性,以COPR、P183和P341这3个站点为<span style="color: black;">科研</span>对象,<span style="color: black;">创立</span>其土壤湿度反演模型,对结果进行分析。其中,COPR站点<span style="color: black;">位置于</span>美国加利福尼亚州维斯塔岛,土壤类型为壤土,站点的东方向有植被遮挡低高度角卫星信号,<span style="color: black;">实验</span>时<span style="color: black;">选择</span>了120°~280°方位角的SNR数据;P183站点<span style="color: black;">位置于</span>美国加利福尼亚州博迪加湾,土壤类型为粗砂壤土,无植被遮挡,但西南方向地形起伏<span style="color: black;">很强</span>,<span style="color: black;">实验</span>时<span style="color: black;">选择</span>了60°~200°和240°~310°方位角的SNR数据;P341站点<span style="color: black;">位置于</span>美国加利福尼亚州惠斯基镇,土壤类型为多石壤土,站点的东北方向植被遮挡较为严重,<span style="color: black;">实验</span>时<span style="color: black;">选择</span>了90°~330°方位角的SNR数据。从各测站有效方位角的SNR序列中提取特征参数,构建PSO-CNN反演模型,粒子群初始参数设置与2.1节一致,将筛选后的特征参数<span style="color: black;">做为</span>模型输入,<span style="color: black;">最后</span><span style="color: black;">实验</span>结果见表 5所示。由表 5<span style="color: black;">能够</span>看出,<span style="color: black;">针对</span><span style="color: black;">区别</span>地表环境的站点本文<span style="color: black;">办法</span><span style="color: black;">拥有</span>较强的适应性,都能有效反演土壤湿度,且各站点的反演结果与P041站点的结果<span style="color: black;">类似</span>,PSO-CNN模型精度最高,BP神经网络模型次之,多星线性和未优化CNN模型稍差。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">表 5</span> COPR、P183和P341站点结果分析<span style="color: black;">Tab. 5</span> Result analysis of COPR, P183 </span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/a52da141efd365838bd05b3b9e6b7fba~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725647938&x-signature=MT1tSX7byp1YP11s97nCQUMW1R4%3D" style="width: 50%; margin-bottom: 20px;"></p>
<h2 style="color: black; text-align: left; margin-bottom: 10px;"><strong style="color: blue;"><span style="color: black;">4 结论与讨论</span></strong></h2>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">高精度、长时序、连续监测土壤湿度变化对生态水文、环境科学、农业生产等<span style="color: black;">行业</span>的<span style="color: black;">科研</span>有着重大现实<span style="color: black;">道理</span>。针对土壤湿度反演的建模问题,本文构建了基于PSO-CNN优化算法的GNSS-IR土壤湿度反演模型,采用多个站点对模型的<span style="color: black;">靠谱</span>性和适应性进行说明和分析。理论分析和<span style="color: black;">实验</span><span style="color: black;">显示</span>。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">(1) 多径干涉特征参数信息能够有效表征本<span style="color: black;">实验</span>区域土壤湿度的变化,但单星反演<span style="color: black;">关联</span>性仅0.586 7,均方根误差为0.038,相对误差约为20%,反演精度<span style="color: black;">不良</span>,其<span style="color: black;">原由</span>是单星反射区域有限、有效时段较短、<span style="color: black;">反常</span>跳变等。将多卫星多频点数据集<span style="color: black;">做为</span>模型网络训练的输入,利用丰富的观测数据,能有效结合<span style="color: black;">区别</span>方位<span style="color: black;">区别</span>时段的卫星反射信号,综合站点地表信息,<span style="color: black;">加强</span>土壤湿度反演精度。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">(2) 针对地表粗糙度、信号受损等非线性<span style="color: black;">原因</span>影响,PSO优化CNN的土壤湿度反演模型能够<span style="color: black;">创立</span>特征参数与土壤湿度之间的<span style="color: black;">繁杂</span>映射关系,其均方根误差为0.015 0,相较于基于单星线性、多星线性、未优化CNN和BP神经网络模型,误差分别降低约60%、27%、31%和21%,能有效<span style="color: black;">控制</span><span style="color: black;">实质</span>环境中下垫面变化、信号损失等<span style="color: black;">原因</span>的影响。与现有<span style="color: black;">办法</span>相比,本文<span style="color: black;">办法</span>能较好<span style="color: black;">控制</span>非线性<span style="color: black;">原因</span>的影响,<span style="color: black;">明显</span><span style="color: black;">加强</span>反演精度,但建模耗时稍长。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">(3) 针对5折交叉验证中第3组<span style="color: black;">实验</span>预测偏高问题,文献中<span style="color: black;">显现</span>类似<span style="color: black;">状况</span>,初步分析其可能<span style="color: black;">原由</span>是该组数据大多来自夏季,杂草生长茂盛,<span style="color: black;">引起</span>土壤湿度的反演值偏大。由图 11第3组<span style="color: black;">实验</span>可知,相较于其他<span style="color: black;">办法</span>,PSO-CNN<span style="color: black;">办法</span>的预测值仅略高于参考值,证明该模型能有效削弱植被影响。后续将围绕这一问题进行深入<span style="color: black;">科研</span>。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">(4) 针对<span style="color: black;">区别</span>下垫面环境,<span style="color: black;">经过</span>分析4个站点<span style="color: black;">实验</span>结果<span style="color: black;">发掘</span>,地表起伏给信噪比时间序列引入<span style="color: black;">更加多</span>噪声,植被覆盖较多<span style="color: black;">引起</span>特征参数<span style="color: black;">包括</span><span style="color: black;">更加多</span>地表植被信息,土壤湿度影响占比减少,3个站点虽然不如P041开阔平坦,但<span style="color: black;">经过</span>筛选<span style="color: black;">靠谱</span>卫星时段,利用有效方位的多源数据,本文<span style="color: black;">办法</span>仍能较好反演土壤湿度,证明了本文<span style="color: black;">办法</span>在较<span style="color: black;">繁杂</span>环境下的适用性。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">综上所述,本文融合多源观测数据,<span style="color: black;">创立</span>粒子群优化卷积神经网络的GNSS-IR土壤湿度反演模型能够有效<span style="color: black;">控制</span>下垫面环境<span style="color: black;">原因</span>的影响,<span style="color: black;">明显</span><span style="color: black;">提高</span>反演精度。进一步深入分析多系统导航卫星数据的高时空分辨率土壤湿度反演精细化模型,是后续<span style="color: black;">科研</span>的<span style="color: black;">重点</span><span style="color: black;">目的</span>和方向。</span></p><span style="color: black;">作者简介</span><strong style="color: blue;"><span style="color: black;"><span style="color: black;">第1</span>作者简介:</span></strong><span style="color: black;">何佳星(1999-), 男, 硕士生, <span style="color: black;">科研</span>方向为GNSS遥感。E-mail: HEJiaxing@cumt.edu.cn</span><strong style="color: blue;"><span style="color: black;">通信作者:</span></strong><span style="color: black;">郑南山, E-mail: </span><span style="color: black;">znshcumt@163.com</span>
<h2 style="color: black; text-align: left; margin-bottom: 10px;"><span style="color: black;">初审:张 琳</span><span style="color: black;">
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">复审:宋启凡</p>
</span><span style="color: black;">终审:金 君</span></h2>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;"><span style="color: black;">新闻</span></strong></p>
</div>
</div>
楼主果然英明!不得不赞美你一下! 百度seo优化论坛 http://www.fok120.com/
页:
[1]