Estimation of SPAD Value in Rice Based on Continuous Wavelet Transform and Back Propagation Neural Network

Hu Wenrui; Gao Qianwen; Yang Huibing; Gao Zhiqiang

doi:10.14083/j.issn.1001-4942.2025.04.019

Abstract：To construct a hyperspectral accurate estimation model of SPAD value of rice leaves,the ex-periment was conducted with the variety Jingliangyouhuazhan as material,and three fertilization treatments were set.The hyperspectral reflectance and SPAD value data were measured continuously and periodically dur-ing the whole reproductive period.The sensitive information of the spectrum was extracted using vegetation in-dex and continuous wavelet transform(CWT),and then the estimation models of SPAD of rice leaves were es-tablished using traditional linear and nonlinear fittingand the algorithm of back propagation neural network(BPNN),whose estimation effects were compared using the coefficient of determination(R2),root mean square error(RMSE),and relative analysis error(RPD).The results showed that the accuracy of the inverse univariate models for SPAD value constructed with nine commonly used vegetation indices based on the tradi-tional methods(linear function,logarithmic function,exponential function,and quadratic function fitting)were lower(RPD＜1.4).Six parent wavelet functions were selected for CWT,which could effectively improve the correlation between leaf hyperspectral reflectance and SPAD value,and the accuracy of the univariate models constructed using the optimal wavelet coefficients of the parent wavelet functions increased obviously,which could reach the level of rough estimation of SPAD value with RPD between1.523 to 1.581.The accuracy of the SPAD estimation models constructed based on the BPNN algorithm were significantly improved compared with the univariate models with RPD between1.832 to 2.342.Among which,the BPNN models constructed with bior3.3 and gaus4 as parent wavelet functions had better estimation capacity with RPD of 2.342 and 2.178,respectively.However,the BPNN model constructed with gaus4 as the parent wavelet function had the phenomenon of overfitting.In summary,the BPNN model constructed using the first 10 optimum wavelet coef-ficients decomposed by bior3.3 had the highest precision with R2 and RMSE as 0.818 and 1.441,respectively,which could make good prediction of SPAD value of rice leaves.This study proved that CWT could effectively extract the sensitive information of spectral characteristics in rice leaves,and the established bior3.3-BPNN model could be used for the monitoring of rice SPAD value,which could provide references for subsequent rapid and nondestructive monitoring of rice SPAD value throughout the whole growth period,and provide the technical support for real-time monitoring of rice growth and development dynamics.