| The yield of apple in China ranks first in the world.The Loess Plateau has become one of the main apple producing areas in the world because of its abundant light resources,unique phenological characteristics and deep soil profile,which provides a superior ecological environment for apple growth.Rapid and accurate acquisition of orchard growth information is the premise of accurate orchard management and high-quality fruit production.Based on the construction of leaf growth information estimation model,this study aims to explore the inversion method of orchard growth information at regional scale and evaluate its application potential.A typical orchard located in Fufeng County,Shaanxi Province was selected as the research object,and the continuous experimental observation was carried out for many years(2016-2018)to explore the method of constructing machine learning models based on characteristic spectra and optimized spectral indices under different spectral processing techniques(resampling,spectral transformation and fractional differential).Support vector machine(SVR)and random forest(RF)were applied to estimate leaf chlorophyll content(LCC),leaf nitrogen balance index(NBI)and leaf water content(LWC)of apple trees at different growth stages.The spectral reflectance of GF-6,Landsat-8(Ls-8)and Sentinel-2(Sn-2)were simulated to estimate the contents of LCC,NBI and LWC in apple leaves at different growth stages.The main research conclusions are as follows:(1)The contents of LCC,NBI and LWC in leaves were 25.45～63.40,6.04～99.76 and39.02～87.09,respectively.The contents of LCC,NBI and LWC in leaves of apple trees increased first and then decreased with the increase of growth stages from the final flowering stage to the harvest maturity stage(1st-5th).Different spectral processing methods changed the shape and detail information of leaf spectral curve.With the increase of resampling interval,the information of spectral bands gradually decreased,and the spectral curves of√R,sin R and tan R were basically similar to those of the original spectrum(R),while the spectral curves of1/R,log R,1/log R and cos R had changed significantly.The spectral details of apple tree leaves could be amplified by fractional derivative spectral transformation.(2)The characteristic spectra after resampling,spectral transformation and fractional differential transformation were significantly correlated with LCC,NBI and LWC of apple leaves.When the resampling interval was 1 nm,the absolute value of the correlation coefficient with LCC,NBI and LWC was the highest.Spectral transformation and fractional differentiation could improve the correlation between characteristic bands and LCC,NBI and LWC.Under the optimal spectral transformation,the absolute values of correlation coefficients between characteristic bands and LCC,NBI and LWC increased by 0.03,0.07 and0.01,respectively.The absolute value of the correlation coefficient between the characteristic band and LCC,NBI and LWC increases by 0.11,0.21 and 0.25 under the best fractional differential order treatment.(3)The estimation models of LCC,NBI and LWC were established based on the characteristic wavebands,and the estimation accuracy of the models was the highest at the resampling interval of 1 nm,and the coefficients of determination(R_c~2)of the SVR models ranged from 0.26 to 0.69,0.19 to 0.71,and 0.35 to 0.46,respectively.The R_c~2 of the RF models ranged from 0.81 to 0.94,0.79 to 0.93,and 0.64 to 0.79,respectively.Spectral transformation could improve the estimation accuracy of the model,and the estimation accuracy of the SVR model of the optimal transformation spectrum could be improved by 0.04,0.07 and 0.03,and the estimation precision of the RF model could be improved by 0.04,0.03 and 0.07.With the increase of fractional differential order,the estimation accuracies of SVR and RF models showd a trend of first increasing and then decreasing.The R_c~2 based on the multivariate SVR model were 0.78,0.76 and 0.62,respectively,while the R_c~2 of the RF model was above 0.85,indicating that the RF model had a high estimation accuracy.The coefficient of determination(R_v~2)of the univariate RF model based on characteristic bands was lower than that of SVR,and the RMSEv was higher than that of SVR,while the accuracy of the multivariate RF model based on characteristic bands was better than SVR.(4)The optimal spectral indices constructed by different spectral processing methods were different at different growth stages.The determination coefficients(R~2)of linear regression between the optimized spectral indices and the contents of LCC,NBI and LWC were the highest under the 1 nm resampling interval,and the R~2 could be increased by 0.18,0.05 and 0.31 compared with the unoptimized spectral indices.Spectral transformation and fractional differentiation could further improved the coefficient of determination.The R~2increased by 0.05,0.05 and 0.02 under the optimal spectral transformation form,and by 0.06,0.14 and 0.09 under the optimal fractional differential treatment.(5)The SVR and RF estimation models of LCC,NBI and LWC were established based on the optimized spectral indices,and the accuracy of the models with 1 nm resampling interval were highest.The R_c~2 of the SVR models ranged from 0.40 to 0.77,0.22 to 0.74,and0.30 to 0.60,respectively.The R_c~2 of the RF model ranged from 0.89 to 0.96,0.84 to 0.96,and0.76 to 0.89,respectively.Spectral transformation and fractional differentiation further improved the estimation accuracy of the model.Under the optimal spectral transformation,the estimation accuracy of SVR model increased by 0.12,0.09 and 0.04,and that of RF model increased by 0.04,0.03 and 0.07.With the increase of fractional differential order,the estimation accuracy of SVR and RF models increased first and then decreased.The R_c~2 of SVR model based on multivariate analysis ranged from 0.51 to 0.83,0.31 to 0.88 and 0.17 to0.79,respectively,while the R_c~2 of RF model was higher than 0.85,indicating that RF model had better estimation accuracy.The R_v~2 of the univariate RF model based on spectral index was lower than that of SVR,and the RMSE of the univariate RF model based on spectral index was higher than that of SVR,while the accuracy of the multivariable RF model was better than SVR.(6)The spectral reflectances of GF-6,Ls-8 and Sn-2 were simulated by the satellite response function values.Based on the band channels with the highest correlation,the SVR and RF estimation models were established.The R_c~2 of the SVR and RF models ranged from0.19 to 0.72 and 0.78 to 0.95,respectively.The optimal spectral index was calculated by using the optimal spectral index equation form at different growth stages.The R~2 of the linear fitting between the spectral index and the leaf LCC,NBI and LWC contents ranged from 0.04 to 0.77.The SVR and RF estimation models of leaf LCC,NBI and LWC contents were established based on the optimized spectral indices,and the RF model had higher modeling accuracy.Compared with the estimation model based on band channels,the estimation model based on optimized spectral indices had higher estimation accuracy.Overall,the estimation accuracy of the model for leaf LCC,NBI and LWC content was the highest at the 3rd stage and the simulated spectra of the three satellites show a trend of Sn-2>GF-6>Ls-8.In conclusion,different spectral processing methods could improve the estimation accuracy of leaf growth information.The estimation model had higher estimation precision in the apple fruit expansion period(3rd).Compared with the univariate estimation model,the multivariate estimation model was more stable.Meanwhile,according to the measured leaf growth parameters and hyperspectral data,the spectral reflectance of GF-6,Ls-8 and Sn-2satellites were simulated,and the SVR and RF estimation models were constructed to retrieve the LCC,NBI and LWC of apple leaves in the region.The estimation model based on Sn-2spectral data had higher estimation accuracy,especially for the RF model based on optimized spectral index,the estimation accuracy of LCC,NBI and LWC of apple leaves were above0.87,0.84 and 0.81,respectively.The results show that the RF model based on optimized spectral indices has a good application potential in the inversion of leaf growth information at regional scale. |
