Comparative study on several hyperspectral estimation models of nitrogen contents in Guanxi honey pomelo leaves

Abstract: 【Objective】Nitrogen(N) content in pomelo leaves is an important index for accurate diagnosis and quantitative evaluation of growth status. Timely and accurate nitrogen diagnosis is an important and key work in precise agriculture. Hyperspectral technology can directly and quantitatively analyze weak spectral differences, which provides a good method for quantitatively analyzing the correlation between plant N contents and spectral parameters. Therefore, the hyperspectral estimation models of N contents in the pomelo leaves were established to provide a basis for rapid, nondestructive and accurate estimation of N content.【Methods】Based on the hyperspectral data of the pomelo leaves and the measured data of N contents, the study firstly analyzed the correlation between the N contents of the pomelo leaves and their original and first-order differential spectra, then analyzed the correlation between spectral characteristic parameters, sensitive band vegetation indices and the nitrogen contents of the pomelo leaves, and found out the spectral parameters with good correlation, finally the partial least squares regression model(PLS), BP neural network regression model (BPNN), random forest regression model(RF) and support vector machine regression model(SVM) of the pomelo leaves were established, and the best estimation model of N contents in pomelo leaves was determined.【Results】There were negative correlation between the original spectral reflectance of the pomelo leaves and the leaf nitrogen contents. The maximum negative correlation coefficients were(-0.683, 569 nm) and(-0.688,704 nm), respectively; there were extremely significant negative correlations between nitrogen and firstorder spectral reflectance in the bands of 440-455 nm, 490-553 nm and 681-705 nm (p < 0.01), among them the maximum negative correlation coefficient was(-0.72, 695 nm), followed by(-0.70, 541 nm);in the bands of 586-627 nm, 633-671 nm and 731-758 nm, it reached extremely significant positive correlation level(p < 0.01), among them the maximum positive correlation coefficient was(0.73, 617nm), followed by(0.72, 753 nm). The sensitive wavelengths of the original spectral curves were 569nm and 704 nm, and the sensitive wavelengths of the first-order differential curve were 541 nm, 617nm, 695 nm and 753 nm. The sensitive wavelengths of 569 nm, 704 nm, 541 nm, 617 nm, 695 nm and753 nm were selected to construct the spectral parameters, and the differential vegetation index(DVI(λ1, λ2)), ratio vegetation index(RVI(λ1, λ2)) and normalized differential vegetation index(NDVI(λ1, λ2))were established. The correlation analysis between the hyperspectral location variables, hyperspectral area variables and hyperspectral vegetation index variables and nitrogen contents of the pomelo leaves showed that most variables had very significant correlation with nitrogen contents of the pomelo leaves.Among the selected three side parameters, only the yellow side parameter Dy(yellow side amplitude)and λy(yellow edge position) did not reach significant correlation, and the red edge parameters and blue edge parameters reached significant correlation. It was found that the correlation coefficients of Db, λr,Rg, SDb, Rg/Rr,(Rg-Rr)/(Rg+ Rr), SDr/SDb,(SDr-SDb)/(SDr+ SDb), DVI'541,753, DVI'695,753, DVI'541,617,DVI'541,695, DVI'617,695, RVI'695,753, NDVI'541,753, NDVI'695,753exceed 0.7, and they had a good ability to estimate the contents of N in the leaves of pomelo, which had reached significant level. The variables with good correlation(r > 0.73)(NDVI'695,753, RVI'695,753, DVI'617,695, DVI'541,617, R'617) in the spectral parameters were selected. The hyperspectral estimation models of nitrogen contents in the pomelo leaves were established using PLS, BPNN, RF, SVM and other methods. The R2(determination coefficient), RMSE(root mean square error) and RE(relative error) of the random forest method were 0.83, 0.97 and3.01%, respectively, and the modeling accuracy was the highest. The R2, RMSE and RE of SVM method were 0.81, 1.02 and 3.04% respectively; the R2, RMSE and RE of BPNN method were 0.80, 1.08and 3.05% respectively; the R2, RMSE and RE of PLS method were 0.75, 1.16 and 3.13% respectively.The R2of the validation models of PLS, BPNN, RF and SVM were 0.79, 0.84, 0.85 and 0.82 respectively, the RMSE were 1.11, 0.94, 0.87 and 0.99 respectively, and the RE were 3.08%, 2.99%, 2.85% and3.03% respectively. The predicted and measured values fit well. Compared with PLS, BPNN and SVM,the RF validation model had higher R2, lower RMSE and lower RE.【Conclusion】Through the comparison of the four hyperspectral estimation models for nitrogen content of the Guanxi honey pomelo leaves, the accuracy of the random forest estimation model was higher than those of PLS, BPNN and SVM. This study would provide a technical basis for monitoring the nitrogen content of the pomelo leaves by spectral remote sensing.