Estimation Of Nitrogen Nutrient Status In Winter Oilseed Rape Based On UAV Multispectral Imagery

Winter oilseed rape is widely grown in China,accounting for more than 90%of the planting area and the rapeseed production in China.Excessive application of nitrogen fertilizer to increase yield is common in planting winter oilseed rape.Therefore,it is essential to quickly,accurately and non-destructively evaluate the nitrogen nutrient status for the precise fertilization.Unmanned Aerial Vehicle（UAV）is a popular remote sensing platform currently because it is flexible and highly maneuverable.However,studies on the estimation of nitrogen nutrient status of winter oilseed rape based on UAV multispectral images are still limited.In this study,8 field experiments with nitrogen fertilizer gradients for winter oilseed rape were carried out in Wuhan,Wuxue and Shayang City,Hubei Province during from2018 to 2022.We measured the canopy spectra with a portable spectroradiometer,acquired multispectral imagery with an UAV,and sampled plants to measure physiological and biochemical parameters.Two quantitative indicators of the nitrogen status in winter oilseed rape,namely leaf nitrogen content(LNC_area,g/m²)and leaf nitrogen concentration(LNC_mass,%),were evaluated using the canopy spectral characteristics and physical/empirical models.For the estimation of LNC_area,a semi-empirical model was established based on the correlation between LNC_area and chlorophyll content（Cab）and the correlation between LNC_area and protein content（Prot）,respectively.The spectrum simulated by the latest version of the PROSAIL-PRO model combined with random forest（RF）regression model were used to build the semi-empirical model to estimate LNC_area.And then the semi-empirical model was applied to UAV multispectral images.For LNC_mass,deep learning models were used to directly extract deep-level features from UAV multispectral images,and empirical models for LNC_mass estimation were built on the deep-level features.The specific conclusions are as follows:（1）In this study,the correlation between Cab and LNC_area of winter oilseed rape was established at the eight-leaf stage,over-wintering stage and budding stage respectively,and the semi-empirical model to invert Cab was built by using the radiative transfer model PROSAIL-PRO and RF regression.Thus,LNC_area was estimated by the correlation with Cab.The results showed that the correlation between Cab and LNC_area was the strongest in the over-wintering and budding stages,with R² of 0.94 and 0.93,respectively,which was significantly higher than the correlation between the two in the eight-leaf stage（R²=0.60）.The semi-empirical model constructed based on the simulation spectrum of PROSAIL-PRO can invert Cab accurately（R²=0.91,RMSE=4.46μg/cm²）.Affected by the correlation between Cab and LNC_area,the estimation accuracy of LNC_area in over-wintering and budding stages were still higher than that in eight-leaf stage.The highest R²value was 0.84 in the over-wintering stage,and the lowest RMSE value was 0.29 g/m² in the budding stage.And then,the above model was extended and applied to the UAV multispectral image,the R² of the inversion result of Cab compared with the true value was 0.82,and the RMSE was as low as 6.03μg/cm².The calculated results of LNC_area at the overwintering stage（R²=0.76,RMSE=0.62 g/m²）and budding stage（R²=0.58,RMSE=0.30 g/m²）were also better than those at the eight-leaf stage（R²=0.45,RMSE=0.40 g/m²）,indicating that the overwintering stage and budding stage may be more suitable for LNC_area estimation by using the semi-empirical inversion model of Cab.（2）The semi-empirical inversion model of Prot was constructed using the spectrum simulated by PROSAIL-PRO combined with the RF regression model,and the LNC_areawas calculated by using the 4.43-fold correspondence between Prot and LNC_area.The results showed that the semi-empirical model constructed based on the simulation spectrum of PROSAIL-PRO performed well,with an R² of 0.89 and an RMSE as low as1.38 g/m².Since the correlation between Prot and LNC_area is relatively direct and less affected by other factors,the estimation accuracy of LNC_area was also higher（R²=0.89,RMSE=0.31 g/m²）.Compared with the estimation of LNC_area based on Cab,the R² was increased by 17.11%,which is a significant improvement,and the RMSE values of the two were almost equal,indicating that the direct inversion of Prot to calculate LNC_area by using the radiative transfer model can achieve more accurate result than that of Cab which was based on the correlation of Cab and LNC_area.（3）Based on UAV multispectral images,four classification models were constructed using convolutional neural network VGG16（Visual Geometry Group 16）and Convolutional Autoencoder（CAE）combined with RF classification model:VGG16,VGG16+RF,CAE+RF and RF（as a comparison）to diagnose nitrogen deficiency in winter oilseed rape.The results showed that using CAE to extract features combined with RF classification model achieved the highest overall classification accuracy of 67.38%and the highest Kappa coefficient of 0.45.（4）An RF model was constructed to estimate LNC_mass using the features extracted by CAE in the nitrogen deficiency diagnosis（CAE+RF）,and compared with VGG16 and RF which used Vegetation Index（VI）as the input feature（VI+RF）.The LNC_mass estimation accuracy obtained by the RF regression model based on the extracted features of CAE was the highest（R²=0.78,RMSE=0.60）,while the VGG16 model（R²=0.72,RMSE=0.69）and VI+RF model（R²=0.70,RMSE=0.62）were slightly lower,indicating that the deep features of UAV images mined by nitrogen deficiency diagnosis are also suitable for the quantitative estimation of LNC_mass.In this study,LNC_area and LNC_mass were estimated based on UAV multispectral images,which provides a feasible solution for the convenient and rapid nitrogen nutrient estimation of winter oilseed rape,and is beneficial to the promotion of precision,refined and high-quality development of rapeseed industry.