基于BiLSTM的猕猴桃根域土壤水分时序反演方法  

Time-series data inversion of soil moisture content in root zone of kiwifruit using BiLSTM

作　　者：李鑫帅 贾泽丰 何景源 高文 潘时佳 牛子杰[1,2] 张东彦 LI Xinshuai;JIA Zefeng;HE Jingyuan;GAO Wen;PAN Shijia;NIU Zijie;ZHANG Dongyan(College of Mechanical and Electronic Engineering,Northwest A&F University,Yangling 712100,China;Shanxi Key Laboratory of Agricultural Information Perception and Intelligent Service Yangling,712100,China)

机构地区：[1]西北农林科技大学机械与电子工程学院,杨凌712100 [2]陕西省农业信息感知与智能服务重点实验室,杨凌712100

出　　处：《农业工程学报》2025年第2期112-119,共8页Transactions of the Chinese Society of Agricultural Engineering

基　　金：国家自然基金联合基金重点项目(U2243235);西北农林科技大学学科建设专项经费项目(Z1011124001)。

摘　　要：根域土壤水分是决定猕猴桃树健康生长与产量的关键因素,尤其在果实膨胀期,土壤水分的动态监测尤为重要。针对传统监测方法无法监测土壤水分持续变化,该研究以眉县猕猴桃实验站为研究区域,采用无人机和地面传感器采集植被光谱反射率及土壤水分数据(共60 d,1440组数据),构建猕猴桃根域土壤含水率的反演模型。通过Pearson和Spearman相关系数筛选了9种植被指数作为模型输入,比较了前馈神经网络(feedforward neural network,FFNN)、长短期记忆网络(long short-term memory,LSTM)及双向长短期记忆网络(bidirectional long short-term memory,BiLSTM)的表现。FFNN由于无法吸收时间序列信息,其在测试集上的表现较差,决定系数为0.269,均方根误差为3.56%。而LSTM和BiLSTM模型利用多日历史数据显著提高预测精度,其中BiLSTM表现最佳,测试集决定系数为0.624,均方根误差为2.45%。研究表明,基于时序模型的土壤水分反演方法可以用于猕猴桃果园果实膨大期的精准监测,也为其他果园作物的水分管理提供一定的理论支持。Soil moisture content in the root zone is one of the most crucial factors in determining the healthy growth and yield of kiwifruit trees,particularly in dynamic monitoring of the soil moisture during fruit expansion.However,traditional monitoring cannot capture the continuous variation in soil moisture.In this study,a time-series data inversion was carried out on the soif moisture content in the root zone of kiwifruit.The fruits were taken from the Meixian kiwifruit experimental station in Baoji City,Shaanxi Province,China.Both unmanned aerial vehicles(UAV)equipped with multispectral sensors and ground-based moisture sensors were utilized to collect the spectral reflectance and soil moisture data over a period of 60 days,respectively,resulting in a total of 1440 datasets.The spectral data was processed to initially extract 20 vegetation indices,such as the Normalized Difference Vegetation Index(NDVI),Soil-Adjusted Vegetation Index(SAVI),and Green Normalized Difference Vegetation Index(GNDVI).The dataset was then refined to identify the most critical features for the soil moisture inversion.Pearson and Spearman correlation coefficients were employed to determine the nine most relevant vegetation indices.The indices were also optimized to reduce the model complexity for the high predictive power.Subsequently,the optimal indices were used as the inputs for three machine learning models:the Feedforward Neural Network(FFNN),the Long Short-Term Memory(LSTM)network,and the Bidirectional Long Short-Term Memory(BiLSTM)network.Among them,the FFNN served as a baseline to compare with the temporal models,due to the temporal independencies in the data.The experimental setup involved training and testing the three models using the dataset.The FFNN model was trained with the input features representing the selected vegetation indices without any temporal information.In contrast,the LSTM and BiLSTM models were designed to utilize the multi-day historical data,thus capturing the temporal dependencies in the soil moisture dynamics

关 键 词：无人机 猕猴桃 土壤含水率 多光谱 遥感 前馈神经网络 长短期记忆网络 

分 类 号：S127[农业科学—农业基础科学] P23[天文地球—摄影测量与遥感]