Retrieving canopy leaf total nitrogen content of winter wheat by continuous wavelet transform

Abstract: The aim of this paper is to monitor the nitrogen nutrition status of winter wheat under stripe rust stress by hyperspectral remote sensing. The experiment was carried out at Beijing Xiaotangshan Precision Agriculture Experimental Base, China (40°10.6′N, 116°16.3′E). The cultivar of winter wheat was Jingdong 8 which was very susceptible to stripe rust. Canopy spectral reflectance data of winter wheat was collected by an ASD Fieldspec FR spectroradiometer and the disease index (DI) was measured through counting the number of wheat leaf under stripe rust stress artificially in the field. Leaf total nitrogen (LTN) content of winter wheat used to calculate DI was measured in the laboratory. The relationship between DI of stripe rust and LTN content of winter wheat was analyzed. The canopy spectra were processed by the method of continuous wavelet transform (CWT) on 10 scales, therefore, a series of wavelet coefficients were obtained in this way. The correlation coefficients between wavelet coefficients and LTN content were calculated, and then, the wavelet coefficients, which had strong correlation with LTN content, were chosen. Several hyperspectral indices were also selected according to previous research results, namely SR (simple ratio index), PRI (photochemical reflectance index), NDVI (normalized difference vegetation index), OSAVI (optimized soil-adjusted vegetation index), SIPI (Structure insensitive pigment index), LIC1 (lichtenthaler index 1), LIC2 (lichtenthaler index 2), LIC3 (lichtenthaler index 3), TVI (triangular vegetation index) and MTVI2 (modified triangular vegetation index 2), which had high correlations with LTN content. Both wavelet coefficients and hyperspectral indices were used as independent variables of models to retrieve LTN content of winter wheat, and support vector machine (SVM) regression method was used to establish the estimation models. The above estimation models of different types of variables were made a comparison. Cross-validation method was used to examine estimation model accuracies of winter wheat canopy LTN content. The test results showed that LTN content of winter wheat decreased gradually when DI of stripe rust increased and the correlation coefficient between LTN content and DI was -0.784 (n=33). The research results indicated that the model prediction accuracies of SR, PRI, LIC2, LIC3, TVI and MTVI2 were very low, the values of R2 were less than 0.700, and root mean square error (RMSE) and relative error (RE) were both high. The estimated models of NDVI, OSAVI, SIPI and LIC1 had higher accuracies than aforementioned hyperspectral indices. However, the accuracies of LTN content estimation models which were built by wavelet coefficients obtained by CWT were generally higher than those of hyperspectral indices. The optimal model was established by wavelet coefficient 423(4) which was acquired by Mexican hat wavelet function, and RMSE and RE were 0.315 and 7.62%, respectively. The wavelet coefficient 423(4) was located in 423 nm where was near to 430 nm which is the nitrogen absorption waveband. The estimation model of wavelet coefficient 663(5) was secondary, which had the prediction RMSE of 0.345 and RE of 8.28%. The variable of 663(5) obtained by Daubechises (Db5) wavelet function was also close to the nitrogen absorption waveband which lies in 660 nm. Therefore, the method of jointing CWT and SVM regression is feasible to retrieve LTN content of winter wheat under stripe rust stress and the estimation models have high accuracies. There is an important practical meaning for preventing diseases of wheat and instructing crop fertilization.