基于无人机高光谱影像和机器学习算法的花生生物量估算方法研究  

Research on estimating peanut biomass using UAV hyperspectral imaging and machine learning algorithm

作　　者：刘涛[1,2] 刘望 杨奉源 张寰 殷冬梅 焦有宙[2] 张梅凤[4] 张全国 LIU Tao;LIU Wang;YANG Fengyuan;ZHANG Huan;YIN Dongmei;JIAO Youzhou;ZHANG Meifeng;ZHANG Quanguo(School of Urban and Rural Planning,Henan University of Economics and Law,Zhengzhou 450046,China;School of Mechanical and Electrical Engineering,Henan Agricultural University,Zhengzhou 450002,China;College of Agronomy,Henan Agricultural University,Zhengzhou 450002,China;Engineering Training Center,Zhengzhou University of Light Industry,Zhengzhou 450103,China)

机构地区：[1]河南财经政法大学城乡规划学院,郑州450046 [2]河南农业大学机电工程学院,郑州450002 [3]河南农业大学农学院,郑州450002 [4]郑州轻工业大学工程训练中心,郑州450103

出　　处：《中国农业大学学报》2025年第3期206-217,共12页Journal of China Agricultural University

基　　金：河南省科技研发计划联合基金青年科学家项目(225200810089);河南省高校科技创新人才支持计划(23HASTIT028);河南省青年人才托举工程项目(2024HYTP018);河南省博士后科研项目(202101041)。

摘　　要：为评估无人机高光谱遥感技术在作物生物量估算应用中的潜力,以荥阳市花生种植试验田作为研究对象,采用无人机搭载高光谱相机收集多个品种花生在成熟期的高光谱影像数据,结合多种机器学习算法,构建花生生物量估算模型,并进行模型的精度评价与对比分析。首先,通过使用Savitzky-Golay滤波器对高光谱影像的反射率进行平滑预处理,并应用Gaussian4小波基函数进行连续小波变换,筛选了53个植被指数作为特征输入;然后,通过皮尔逊相关系数法进行敏感植被指数筛选,并利用筛选出的植被指数分别构建支持向量回归(Support Vector Regression,SVR)、随机森林(Random Forest,RF)、卷积神经网络(Convolutional Neural Networks,CNN)、粒子群优化(Particle Swarm Optimization)的SVR、粒子群优化的RF、粒子群优化的CNN等花生生物量估算模型,后进行模型精度评价。结果表明:深度学习模型CNN相比于传统的机器学习模型如RF和SVR等,在花生生物量的预测精度上表现更优;CNN模型在测试集上的决定系数(R2)为0.710,RMSE为0.371 kg/m2,MSE和MAE分别为0.138和0.329 kg/m2;通过粒子群算法PSO进行参数优化后,RF、SVR、CNN模型的预测精度都有提升,其中CNN的提升较为明显,决定系数(R2)提升约为8.2%。因此,在花生的收获期使用PSO对CNN参数优化后的模型对于花生整体生物量的估算最为准确。本研究可为精确预测花生生物量提供科学方法,为智慧乡村建设提供有力支撑。Accurate and efficient crop biomass estimation is of remarkable importance to identify superior varieties,regional production management and food security assessment.In order to evaluate the potential of UAV hyperspectral remote sensing technology in crop biomass estimation,a peanut planting test field in Xingyang City was taken as the research object.The hyperspectral image data of multiple varieties of peanuts at maturity were collected by UAV equipped with hyperspectral cameras.A peanut biomass estimation model was constructed by combining various machine learning algorithms.The accuracy of the model was evaluated and compared.Firstly,the reflectance of the hyperspectral images was smoothed and preprocessed using the S-G filter,and continuous wavelet transform was applied using the Gaussian4 wavelet basis function to screen 53 vegetation indices as feature inputs.Secondly,the sensitive vegetation indexes were selected through the Pearson correlation coefficient method.Lastly,the identified vegetation index was used to construct Support Vector Regression(SVR),Random Forest(RF),Convolutional Neural Network(CNN),Particle Swarm Optimization Support Vector Machine(PSO-SVR),Particle Swarm Optimization Random Forest(PSO-RF),and Particle Swarm Optimization Convolutional Neural Network(PSO-CNN)models for estimating peanut biomass were constructed and their accuracy was evaluated.The results indicated that:The deep learning model CNN demonstrated a superior performance in the prediction accuracy of peanut biomass compared to traditional machine learning models such as RF and SVM.The CNN model had a determination coefficient(R2)of 0.710,RMSE of 0.371 kg/m2,MSE of 0.138 kg/m2,and MAE of 0.329 kg/m2 on the test set.After parameter optimization using the Particle Swarm Optimization(PSO)algorithm,the prediction accuracies of RF,SVR,and CNN models were improved,with the CNN model showing the most significant improvement,with an increase of approximately 8.2%in the determination coefficient.Therefore,using the PSO-optimized CN

关 键 词：高光谱影像 连续小波变换 花生生物量 机器学习 智慧乡村 

分 类 号：S107[农业科学—农业基础科学]