基于多光谱成像和卷积神经网络的玉米作物营养状况识别方法研究

Recognition method for corn nutrient based on multispectral image and convolutional neural network

水肥一体化自动装备的使用能够有效提高水肥资源利用率,但需要在作业前获知作物的营养状况及水肥需求量,而通过人工手持测量仪器来获取这些信息,存在着时效性差和劳动强度大等缺点。针对以上问题,本研究以常见的作物玉米为研究对象,使用大疆精灵Ⅲ无人机携带RedEdge-M多光谱相机在田间上空采集玉米多光谱图像,同时使用YLS-D系列植株营养测定仪测量玉米植株的氮素和水分含量等营养信息,根据这些信息将采集的图像分为3个等级(每个等级共包含530幅五通道图像,其中480幅作为训练集,50幅作为验证集),提出了一种基于卷积神经网络的玉米作物营养状况识别方法。并基于TensorFlow深度学习框架搭建了ResNet18卷积神经网络模型,通过向模型输入彩色图像数据和五通道多光谱图像数据,分别训练出适合于彩色图像和多光谱图像的玉米植株营养状况等级识别模型。试验结果表明:训练后的模型能够识别玉米作物的彩色图像和多光谱图像,能够输出玉米的营养状况等级和GPS信息,识别彩色图像模型在验证集的正确率为84.7%,识别多光谱图像模型在验证集的正确率为90.5%,模型训练平均时间为4.5h,五通道图像识别平均用时为3.56s。该识别方法可快速无损地获取玉米作物的营养状况,为有效提高水肥资源利用率提供了方法和依据。

Excessive application of water and fertilizer not only causes resources serious waste of,but also causes serious environmental pollution.The implementation of precision irrigation and fertilization can effectively reduce nutrient loss and environmental pollution,save irrigation water and improve the utilization rate of water and fertilizer resources,which is one of the important ways to promote the sustainable development of agriculture.The use of the integrated water-fertilizer equipment can effectively improve the utilization rate of water-fertilizer resources,but it is necessary to know the nutritional status of crops and water-fertilizer demand before operation.To acquire the information by hand-held measuring instruments,there are some disadvantages,such as poor timeliness and high labor intensity.In response to the above problems,this study took the common corn crop as an example,used the DJI Phantom III drone to carry RedEdge-M multispectral camera to collect multispectral images of corn crops over the fields,and measured nitrogen and moisture content of corn plants by YLS-D series plant nutrition tester.Based on this information,the collected images were divided into 3 levels,each level contains 530 five channel images(2650 single channel images),including 480 five channel images(2400 single channel images)in the training set and 50 five channel images(250 single channel images)in the verification set,and a method of identifying the nutritional status of corn crops based on convolutional neural network was proposed.Based on the TensorFlow deep learning framework,ResNet18 convolution neural network model was constructed.By entering color image data and five-channel multispectral image data into the model,the nutritional status recognition model of corn plant suitable for color image and multispectral image was trained,and the experimental results showed that the trained model could be used to recognize the multispectral images of corn,and the nutritional status of corn,topdressing guidance and GPS information could be outputted,the correct rate of the recognition color image model in the verification set was 84.7%.The correct rate of identifying multispectral image model in the verification set was 90.5%,the average time of model training was 4.5h,and the average time of recognizing a five channel image is 3.56 seconds,which can detect the nutritional status of corn crops quickly and undamaged,and provides a theoretical and technical basis for the accuracy of the application of water fertilizer in intelligent agriculture.