Diagnosis Of Nitrogen Nutrition In Winter Rape Based On UAV Multispectral Image

Real-time and accurate diagnosis of nitrogen nutrition is one of the important measures to achieve precise fertilization and high yield of crops.The advantages of UAV's portability and flexibility make it a research hotspot of crop quantitative remote sensing in recent years.Research on nitrogen nutrition diagnosis of winter oilseed rape using multi-spectral images of UAV is still relatively limited.Previous studies have shown that adding texture metrics to improve a winter oilseed rape nitrogen nutrition estimation model based solely on the vegetation indexes(VIs)can better reflect the spatial characteristics of the image.Based on this,taking winter oilseed rape in the overwintering period as the research object,taking the nitrogen nutrition characteristics as the starting point,and using the field plot nitrogen fertilizer test as the basis,comprehensively using the multi-year(2016-2019)winter oilseed rape data,multiple test sites(Wuxue,Sha Yang),carried out drone flight and image processing in the critical growth period of winter rape,combined with analysis of physiological and biochemical parameters,from the four aspects of plant nitrogen concentration,above-ground biomass,nitrogen nutrition index of overwintering period,and yield estimation Modeling and exploration,the following conclusions have been obtained:(1)In the estimation of the nitrogen concentration(PNC)of winter rapeseed plants using two experimental fields in two regions in different regions,three modeling methods of general linear regression,partial least squares regression(PLSR)and random forest(RF)regression were used.Linear regression is used for factor correlation screening.In the partial least squares regression,variable projection importance index(vip)screening factors are used.In random forest regression,factor importance ranking screening factors are used.The combined model of texture metrics and vegetation index Compare with the multiple statistical regression model of simple vegetation index.The results show that in single exponential linear regression,the red edge chlorophyll index(CIred edge)and anthocyanin reflectance index(ARI)PNC estimation model R~2 is the highest(CIred edge:R~2=0.70,RMSE=0.64;ARI:R~2=0.70,RMSE=0.65)but the RMSE error is large.In multiple statistical regression,the combined RF regression model of vegetation index and texture metrics can better estimate PNC(R~2=0.97,RMSE=0.51).Although the model has overfitting phenomenon,the RMSE value of the combined model has dropped by 21.5%.There is a significant reduction.Therefore,the addition of texture metrics has an effect on improving the accuracy of the model,but according to the results of this experiment,it is not possible to recommend a texture metric that has a significant advantage for PNC estimation.(2)In the estimation and verification of the above-ground biomass of winter oilseed rape using three experimental fields in two years in different regions,three methods of general linear regression,PLSR and RF regression were used,and the factors were carried out by general linear regression.Correlation screening,and also compared the combination model of texture metrics and VIs with the multiple statistical regression model of simple VI.The result is that in the single index linear regression,RVI andedge have higher R~2,but the RMSE PLSR modeling,the PLSR combination model of VIs and texture metrics provides more accurate AGB estimates,and is based compared with VIs'PLSR model,the RMSE value of the verification data set is reduced by 7.3%,and compared with the linear model,it is reduced by 7.9%;in RF modeling,the RF regression combination model of VIs and texture metrics provides more accurate AGB.It is estimated that the RMSE value of the validation data set is reduced by 15.7%compared to the accuracy of the RF model based on VIs only,and the RMSE value is reduced compared to the accuracy of the linear regression model based onedge 11.1%.However,the RF model also overestimates the AGB of the validation data set.Among the three methods,the RF combined model based on VIs and texture metrics has the highest estimation accuracy.The addition of texture metrics has improved the accuracy of the model.According to the evaluation of the important variables,both methods selected CIred edge and RVI as the most important input variables,and selected NDVI contrast as one of the important texture metrics.However,this result is highly dependent on the data set,and it is currently not possible to recommend the VI and texture metrics that are generally suitable for estimating AGB.(3)Based on the above PNC and AGB estimation models,the nitrogen dilution curve of winter rapeseed obtained from previous studies was used to obtain the critical nitrogen concentration,and then the NNI model of winter rapeseed was constructed.Based on the relative yield,the NNI threshold was determined to be 1.15.Diagnose the nitrogen nutrition status of the field.The results show that the NNI values of N0,N4 and N6 are lower than the critical value and are nitrogen-deficient field blocks;the NNI values of N8,N12 and N16 are approximately equal to the critical value.According to the actual fertilization experience,N8,N12 and N16 can be diagnosed as nitrogen fertilizer Suitable for fields.This result is consistent with the previous research results of the research group,but the accuracy of judgment also needs to quantitatively calculate the amount of top dressing,and judge based on the effect of top dressing.(4)In the estimation of winter oilseed rape production using four key growth periods in three plots over a two-year period,the three VIs of NDVI,SAVI andedge,NIRv and NDYI were used.First,a single index for each key growth period was conducted with the regression of yield,the NIRv of flowering period provides the optimal yield estimate(training time-integrated VI was used to estimate yield during the growing season.The total biomass of NIRv was found to provide the most accurate prediction of rapeseed yield for training and validation data sets(training set:RMSE=534.60,validation set:RMSE=504.02).Since serious nitrogen deficiency rarely occurs in the Yangtze River Basin,the accuracy of yield prediction for treatments with a nitrogen fertilizer dosage higher than 120 kg/hm~2 was specifically evaluated.The RMSE of the predicted yield value,the NIRv measured at flowering stage provides the most accurate estimate compared to the VIs of other studies.Therefore,it is recommended that NIRv be used as an index to predict yield,which is more stable throughout the growing period and shows better estimation ability during the flowering period.