Study On Cotton Nitrogen Nutrition Monitoring Based On "UAV-Ground" Spectral Fusion

Objective:Nitrogen fertilizer is the primary fertilizer for improving cotton yield and quality,and the environmental pollution caused by improper application of nitrogen fertilizer is becoming increasingly serious.Clarifying the nitrogen demand patterns of crops and improving nitrogen fertilizer utilization efficiency are of great significance for precise agricultural management and ecological environment protection.As a supplementary means of ground and satellite remote sensing,unmanned aerial vehicle remote sensing technology has the advantages of strong flexibility,high affordability,and large-scale monitoring.It is one of the necessary technologies for large-scale farmland information monitoring.This study is based on UAV hyperspectral technology to explore the impact of UAV flight altitude on cotton nitrogen nutrition monitoring.It clarifies the cotton feature spectrum extraction technology based on UAV hyperspectral images and integrates it with ground spectra to construct a nitrogen content monitoring model,providing technical support for improving the applicability and stability of cotton nitrogen content monitoring modelsMethods:This study was conducted in 2019 and 2020 at the experimental field of Shihezi University in Shihezi,Xinjiang Uygur Autonomous Region.Six nitrogen fertilizer treatments were set up in the experiment,including N₀(0 kg·ha^-1),N₁₂₀(120 kg·ha^-1),N₂₄₀(240 kg·ha^-1),N₃₃₀(330 kg·ha^-1),N₃₆₀(360kg·ha^-1)and N₄₈₀(480 kg·ha^-1),of which N₃₃₀ is the local conventional fertilizer amount.The leaf nitrogen content（LNC）of cotton was obtained at the early flowering stage,the full flowering stage,the full boll stage,and the boll-opening stage of cotton,and the hyperspectral images of unmanned aerial vehicles with flight heights of 60 m,80 m and 100 m were synchronously obtained.Spectral angle mapper（SAM）and iterative self-orgnizing data analysize technique（ISODATA）were used to classify the image pixels and reduce the interference of soil and cotton leaf shadows on the feature spectra of cotton.The correlation between nitrogen and image pixels spectra determines the necessity of classification.Based on Pearson correlation coefficient（PCC）and randomized sparse models（RSM）,the feature wavelengths and sensitivity indices of cotton nitrogen were selected,and the importance scores of feature wavelengths and sensitivity indices on nitrogen were calculated using random forest.The monitoring model of nitrogen in cotton was constructed based on multiple linear regression（MLR）,partial least squares regression（PLSR）,random forest（RF）,and light gradient boosting machine（LGBM）.The accuracy of the model was evaluated using three evaluation indexes,namely,coefficient of determination（R²）,root mean square error（RMSE）and relative error（MAE）.Construction and validation of a nitrogen monitoring model for UAV and ground-scale spectral fusion data using a feature-level fusion approach.Results:（1）The influence of different flight altitude of UAV on LNC feature wavelength and sensitivity index is analyzed.The results showed that spectral reflectance,vegetation index and cotton LNC were significantly correlated at a flight altitude of 60 m.The feature wavelengths of nitrogen were located from 520～670 nm and 850～1000 nm.The band reflectance of 560 nm（green light）and 930～980 nm had the highest importance score of 75%for nitrogen,and VOGI,GMI,RM,NDVI₈₁₀ had the highest importance score of 56%for nitrogen.At the flight altitude of 80 m,the 578 nm,710～770 and 930～980nm bands had the highest importance score for nitrogen with 69%,VOGI,m ND₇₀₅,SAVI,MACI had the highest importance score for nitrogen with 59%.At flight altitude of 100 m,565 nm,940～990 nm bands had the highest importance score for nitrogen with was 61%,and ZMI,RVI,R/G,GMI,and MTCI had the highest importance score of 71%for nitrogen,while the importance scores of other feature wavelengths and sensitivity indices for nitrogen at all altitudes were less than 10%.（2）The research of cotton feature spectral information extraction based on UAV hyperspectral pixel classification was explored.The results showed that SAM and ISODATA classification reduced the interference of other image pixels such as soil and leaf shading to the feature spectral information of cotton.The PCC of SAM reflectance,index and nitrogen were higher than ISODATA,in which,the PCC was higher than 0.60 from 526～650 nm,and the highest PCC was 0.76 at 540 nm.Using PCC and RSM to screen the feature wavelengths,the wavelengths with higher importance scores for nitrogen were distributed from 530～560 nm and 900～1000 nm.After BPC,SAM,and ISODATA,R/G,GRVI,GRVI,VOGI,GRVI,and m ND₇₀₅ had higher importance scores for nitrogen with 0.21,0.49,0.17,0.49,0.28,and0.22,respectively.This study shows that the importance scores of individual bands and indices for nitrogen are lower（<0.05）.The PCC of nitrogen and image pixel spectra of SAM is significantly higher than that of ISODATA.（3）A multi temporal LNC monitoring model for cotton was constructed based on UAV hyperspectral data.The results showed that the feature wavelengths screened based on PCC and RSM are mainly located in the green light region,the red edge position,and the near-infrared region.The index count with significant correlation with nitrogen content was as follows:at the full flowering stage>the full boll stage>the early flowering stage>the boll-opening stage.At each growth stage,GRVI,RVI,MACI,GMI,and m SR₇₀₅ were significantly correlated with LNC.Based on unmanned aerial vehicle hyperspectral technology,precise monitoring of nitrogen nutrition at various growth stages of cotton can be achieved.The monitoring accuracy of the model is as follows:the full boll stage>the full flowering stage>the early flowering stage>the boll-opening stage.The accuracy of different modeling methods was LGBM>RF>MLR>PLSR,among which R²c and R²v of LGBM and RF reached above 90%,RMSEc and RMSEv were not higher than 1g·kg^-1,MAEc and MAEv were less than 1.MLR and PLSR were less effective monitoring.（4）The cotton LNC monitoring model is constructed and verified based on the"UAV-ground"spectral fusion.The results showed that compared with the model built by the ground spectrum,the"UAV-ground"spectral fusion model R² increased by 0.02～0.05,and RMSE and MAE decreased by 0.1～0.3.Compared with the model built by the UAV hyperspectral,R² increased by 0.2,and RMSE and MAE decreased by0.2～0.5.The accuracy of the models in each growth stage is high,among which the LGBM model constructed by the fusion of the"UAV-ground"sensitivity index has the validation accuracy of R²v,RMSEv and MAEv of 0.95,0.80 g·kg^-1 and 0.29 at the full boll stage of cotton.The applicability and stability of the model are significantly improved through the"UAV-ground"spectral fusion data.Conclusions:To sum up,hyperspectral data of UAV at different flight heights have higher importance scores for LNC,and the flight altitude of 60 m has greater potential for LNC monitoring.Using the 60 m hyperspectral image of UAV and pixel classification technology（SAM,ISODATA）can effectively reduce the impact of soil and leaf shadow on cotton characteristic spectral information.Using the classification method of SAM,combined with"UAV-ground"spectral fusion,can improve the precision of cotton multi temporal nitrogen monitoring（LGBM,RF accuracy up to 90%）,providing theoretical basis and technical support for cotton nitrogen nutrition monitoring technology innovation and optimization,monitoring accuracy and stability improvement.