| Nitrogen is an important nutrient in the growth and development of crops.Therefore,real-time detection and evaluation of nitrogen content of crops are of great significance for crop nitrogen nutrition testing,field precision management and crop growth prediction.At present,the nitrogen nutrition diagnosis methods of crops generally use field destructive sampling and indoor analysis of crop tissue.Although the indoor experiment results are more accurate,the method consumes a lot of manpower,material resources and hysteresis.Therefore,the current nitrogen nutrition diagnosis method is difficult to achieve real-time,rapid and non-destructive diagnostic requirements.However,with the rapid development of UAV remote sensing technology in China,it provides a new diagnostic method and technical means for crop nitrogen nutrition diagnosis.In this study,the UAV remote sensing platform was used to collect the canopy hyperspectral image of japonica rice,and the different classification methods were verified and analyzed to extract the pure canopy canopy spectral data.And the canopy hyperspectral data and canopy leaf nitrogen content of japonica rice at different growth stages under three different nitrogen fertilizer gradient levels were taken as data sources to compare and analyze the different spectral pretreatment methods.Meanwhile,the characteristic bands and vegetation index of canopy leaf nitrogen content of japonica rice were extracted.Four inversion models of canopy leaf nitrogen content of japonica rice were constructed to realize the accurate diagnosis of the nitrogen content in regional range of northeastern japonica rice.The specific research results are as follows:(1)This study is based on the UAV hyperspectral remote sensing platform to collect hyperspectral remote sensing images of japonica rice canopy at different growth stages.The minimum distance method,maximum likelihood method and support vector machine are used to classify the remote sensing images and extract the pure canopy canopy spectral information.The comparative analysis shows that the classification of hyperspectral image of japonica canopy is best by using support vector machine classification,the classification accuracy is89.1895%,and the Kappa coefficient is 0.7925.(2)In this study,1-Der,logarithm,SNV and SG smoothing method were used to preprocess the original spectral data,and construct a partial least squares regression model for comparative analysis.The results show that the SNV has better denoising effect.The R~2 and RMSE of the training set and the verification set are:0.5040,0.7788,and 0.5399,0.3662.Obviously better than other pretreatment methods,it can effectively eliminate or reduce the influence of external environmental factors such as illumination etc.and improve the efficiency and accuracy of modeling.(3)Based on the pre-processed spectral data,this study used UVE and CARS to screen the characteristic bands of the full spectral data at different growth stages.The number of characteristic bands under different growth periods are 9 in the tillering period,7 in the jointing stage,and 6 in the heading period,which significantly reduces the data dimension and eliminates redundant information.At the same time,the best normalization,ratio and difference vegetation index of different growth stages were screened out by using the determinant coefficient R~2 isopotential map.(4)In this study,multiple linear regression,BP neural network,RBF neural network and self-adaptive differential evolution extreme learning machine(Sa De-ELM)were used to model the nitrogen content of japonica rice leaves at different growth stages.The results of model inversion show that the best estimate results are obtained by using self-adaptive differential evolution extreme learning machine,whether multi-feature band combination is used as model input or multi-vegetation index combination as model input. |
PDF Full Text Request