基于深度学习的大棚及地膜农田无人机航拍监测方法

摘要

With the development of precision agriculture,the demand on rapidly obtaining the area and geographical distribution of greenhouses,plastic-mulched landcover is increased.However,using the interpretation method for satellite remote sensing images to process unmanned aerial vehicle (UAV) images is not ideal,due to the complex feature extraction,low recognition accuracy,long processing time and so on.To circumvent this issue,a UAV aerial monitoring method was proposed based on deep learning for greenhouses and plastic-mulched landcover monitoring.The six-rotor UAV equipped with Sony NEX-5k camera captured aerial photographs in the Wangyefu town of Chifeng City.The 558 UAV images were orthographically corrected and stitched.The five fully convolutional network (FCN) variants,i.e.the FCN-32s,FCN-16s,FCN-8s,FCN-4s and FCN-2s models were built by multi-scale fusion.The modes were trained end-to-end by the stochastic gradient descent algorithm with momentum.The features were extracted and learned from the photographs automatically.The FCN models were compared with two economic softwares,i.e.the pixel-based classification method of ENVI and the object-oriented classification method of eCognition.The results showed that the FCN-4s was the best model on the identification of greenhouses and plastic-mulched landcover.The average overall accuracy of test area was 97％,while that of pixel-based classification method and the object-oriented classification method was 74.1％ and 81.78％,respectively.The average runtime of the FCN-4s was 16.85 s,which was 0.06％ and 5.62％ of those of pixel-based classification method and the objectoriented classification method,respectively.The proposed method demonstrated high recognition accuracy and fast speed,which can meet the demand on UAV monitoring of facilities agriculture.%随着精准农业技术的发展,快速获取大棚和地膜农田面积及地理分布的需求越来越大,但沿用面向卫星遥感影像的解译方法处理无人机航拍影像,存在特征选择复杂、识别精度较低、处理时间长等问题.基于此,本文提出一种基于深度学习的大棚及地膜农田无人机航拍监测方法,即采用六旋翼无人机搭载索尼NEX-5k相机进行航拍作业,对采集到的558幅赤峰市王爷府镇地区的无人机航片进行正射校正与拼接,构建全卷积神经网络(Fully convolutional network,FCN),通过多尺度融合的方法实现了FCN的5个变种模型:FCN-32s、FCN-16s、FCN-8s、FCN-4s、FCN-2s,使用带动量的随机梯度下降算法端到端训练模型,自动提取并分类影像特征.FCN模型与ENVI商用遥感软件的基于像素的分类方法、eCognition软件的面向对象的分类方法对比后表明:FCN-4s模型为识别大棚和地膜农田的最佳模型,对于测试区域的平均整体正确率为97％,而基于像素的分类方法平均整体正确率为74.1％,面向对象的分类方法平均整体正确率为81.78％.FCN-4s模型平均运行时间为16.85 s,是基于像素的分类方法运行时间的0.06％,是面向对象的分类方法运行时间的5.62％.本方法可快速准确获取大棚和地膜农田的地理分布及面积,满足设施农业对无人机航拍监测的需求.