Inversion Of Nitrogen Content In Apple Tree Canopy Based On Shadow Removal From UAV Multispectral/Hyperspectral Images

Nitrogen is an indispensable nutrient element in the growth and development of fruit trees,which has an important impact on the yield and quality of fruit trees.The traditional measurement method of nitrogen content is Kjeldahl method.Although this method has high accuracy,it is difficult to meet the needs of field real-time monitoring.UAV can carry simultaneous interpreting of images with high temporal and spatial resolution with different sensors.It has been widely applied in the research fields of vegetation nutrient diagnosis.Because the observation direction of UAV sensor is not consistent with the direction of direct sunlight,there are shadows in UAV remote sensing images,which weaken the canopy spectral information and reduce the retrieval accuracy of canopy nitrogen content.Therefore,how to remove the shadow in the canopy remote sensing image of fruit trees to improve the retrieval accuracy of nitrogen content of fruit trees is an urgent practical problem.In this study,the apple orchard in Qixia City of Shandong Province was taken as the research area,and the canopy of apple trees was taken as the research object.Based on the UAV carrying multispectral and hyperspectral sensors,the canopy images of different tree types were obtained.Based on the UAV multispectral and hyperspectral images and the data of Apple canopy nitrogen content measured in the laboratory,the Normalized Difference Canopy Shadow Index(NDCSI)was used to remove the shadow of the fruit tree canopy in the remote sensing image and extract the spectral information of the fruit tree canopy.Simultaneous interpreting the difference of spectral characteristics between different tree types and different sensor canopy of fruit trees before and after shadow removal,the influence of shadow on spectral information of canopy of different tree types and different sensor fruit trees was analyzed.Based on correlation coefficient method and improved correlation coefficient method(Modified Correlation Coefficient),the simultaneous interpreting method was applied to detect the canopy spectral information.The sensitive wavelengths of multispectral and hyperspectral images of fruit tree canopy were selected by using the method(MCCM),the spectral parameters were constructed,and the linear and nonlinear nitrogen content inversion models of apple tree canopy were established.The main results are as follows:(1)Based on NDCSI,the shadow of UAV remote sensing image of fruit tree canopy was effectively removed.Based on NDCSI setting reasonable threshold,in UAV multispectral image,the reasonable threshold of shadow removal was 0.35 for arboroid fruit tree image and 0.4 for dwarfing fruit tree image;in UAV hyperspectral image,the reasonable threshold of shadow removal was 0.035 for arboroid fruit tree image and 0.04 for dwarfing fruit tree image,which realizes the effective shadow removal of UAV image of fruit tree canopy.(2)Simultaneous interpreting the influence of shadow on spectral data of canopy and canopy of different sensor and tree types.After removing the shadow,the canopy spectral reflectance values of the two tree types were improved;the spectral images of the two tree types and the two sensors were obviously affected by the shadow in the red edge and near infrared band;the canopy spectral reflectance values of arboroid trees were lower than that of dwarfing trees due to the influence of the arrangement and canopy structure of fruit trees.(3)Simultaneous interpreting the sensitive wavelength and spectral characteristic parameters of nitrogen content in different sensor data of UAV.Based on the correlation coefficient method,the sensitive bands of nitrogen content in Apple canopy multispectral data were selected,and the sensitive bands were green band and red band;based on the improved correlation coefficient method,the sensitive wavelengths of nitrogen content in Apple canopy hyperspectral data were selected,and the sensitive wavelengths were 470nm?474nm?490nm?514nm?582nm?634nm?682nm.RVI_(NIR/R)?RVI_(NIR/G)?RVI_{(Red-edge/NIR)}?RDVI?DVI?NDVI?GNDVI?GRVI?NRI?OSAVI?NDCI?TVI?R-M?MSAVI?MSRVI?N-L VI were constructed based on the multi-spectral data of apple canopy before and after shadow removal 16 spectral parameters of vegetation index and 27 non-vegetation index spectral parameters of G?R?E?N?Ln G?Ln R?Ln E?Ln N?1/G?1/R?1/E?1/N?(?)?(?)?(?)?(?)?G×R?G×E?G×N?R×E?R×N?E×N?G×R×E?G×R×N?G×E×N?R×E×N?G×R×E×N were used for correlation analysis.It was found that the correlation between spectral information and sensitive spectral parameters of nitrogen content before shadow removal was lower than 0.35,while the correlation between spectral parameters selected after shadow removal was higher,and the final selected sensitive spectral parameters were Green?(?)?ln G?1/G?G*R?G*R*N,and the correlation coefficients were-0.681,-0.681,-0.679,0.676,-0.672,-0.664,respectively.The sensitive bands selected from canopy hyperspectral data before and after shadow removal were randomly combined to construct four spectral parameters:DSI_(x,y)?RSI_(x,y)?NDSI_(x,y)and DDI_(x,y,z).The results showed that the correlation between spectral parameters and nitrogen content based on three bands was higher than that based on two bands.The final sensitive spectral parameters were DDI_{(474,514,634)}?DDI_{(474,582,634)}?DDI_{(490,582,634)}?DDI_{(470,582,634)}?DDI_{(514,582,634)}?DDI_{(474,490,634)},and the correlation coefficients were 0.697,0.735,0.728,0.718,0.716,0.701,respectively.(4)The inversion model of the best nitrogen content of apple tree was constructedThe accuracy of linear model(multiple linear regression,partial least squares model)and nonlinear model(support vector machine,BP neural network model)based on hyperspectral image after shadow removal was higher than that based on multispectral image,and the accuracy of nonlinear model was higher than that of linear model.Among them,the optimal model was the support vector machine model based on the hyperspectral image after shadow removal,the model set R²was 0.733,RMSE was 6.00%,n RMSE was 12.76%;the verification set R²was 0.671,RMSE was 4.73%,n RMSE was 14.83%.