Diagnosis and recognition of potassium stress degree in rice based on PND-Net model
[Objective]In order to quickly and accurately diagnose the degree of potassium stress in rice,the application of deep learning technology was promoted in the nutritional diagnosis of rice.[Method]The leaf images of rice cultivated under different degrees of potassium fertilization stress were obtained by flatbed scanner as the research object,and the data set was expanded by random Angle rotation,horizontal inversion,contrast enhancement,sharpening enhancement,mixing enhancement,Gaussian noise,pepper and salt noise,multiplicative noise and other methods.A Potassium Nutrition Diagnosis Network(PND-Net),a deep learning diagnostic model for potassium stress in rice,was proposed to accurately identify potassium stress categories in rice.PND-Net integrated two multi-scale feature extraction modules,MS(multi-scale)and RMS(Residual multi-scale),to extract the features sensitive to potassium stress response from different scales of images.Coordinate Attention module was introduced to examine the direction and position information in the leaf image,thus improving the ability of the model to capture information related to potassium stress.Finally,the Multi-Branch Inverted residual module is introduced to enhance the capability of long distance feature capture.[Result]Compared with DenseNet-201,ResNet-34 and GoogLeNet-V3 models,PND-Net had better performance in recognizing the degree of potassium stress in rice,with the highest recognition accuracy of 77.11%and 84.31%at tillering stage and jointing stage.In addition,PND-Net also achieved remarkable results in Plant Village,which further verified the generalization ability of the model.The results of model evaluation further confirmed that the PGN-NET had excellent recall rate and precision,77.23%and 77.14%at tillering stage and 84.4%and 84.33%at jointing stage,respectively.[Conclusion]PND-Net can accurately identify the degree of potassium stress in rice.It provides a theoretical basis for timely topdressing of rice at each growth stage,and a new feasible method for nutritional diagnosis and recognition of rice and other crops.
ricedeep learningdiagnosis of potassium stress degreecoordinate attentionmultiscale feature
万方数据
维普
