Study On Estimation Of Physiological And Biochemical Parameters Of Kiwifruit Based On Stacking Ensemble Learning

The kiwi industry in Shaanxi Province has developed rapidly in scale and has become a characteristic industry to improve the quality of local economic development and increase farmers’income.However,there are still problems in nutrient management of kiwi orchards,such as excessive fertilization and outdated agricultural techniques,which affect the quality and economic benefits of kiwi.Therefore,the growth status of kiwifruit is monitored by remote sensing technology to clarify the spatial distribution of kiwifruit orchards,which is of great significance for timely adjustment and optimization of kiwifruit cultivation management.This study focuses on the leaves of kiwifruit in the northern foothills of the Qinling Mountains from May to August 2022.The hyperspectral reflectance,leaf chlorophyll content(LCC),leaf anthocyanin content(Anth),leaf Nitrogen balance index(NBI)were measured respectively to obtain high-resolution RGB images in the study area,and the original spectra were smoothed and mathematically transformed.By analyzing the correlation between the physiological and biochemical parameters of kiwifruit in different growth stages and the hyperspectral parameters in the range of 380～1000 nm,the input characteristics of the model were screened,the feature recursive elimination(RFE)method was used to screen the best modeling subset in each growth stage,and the lightweight gradient elevator(Light GBM),random forest(RF),K-nearest neighbor(KNN)and adaptive enhancement(Adaboost)algorithm were selected as the base model,Ordinary linear regression(LL)and ridge regression(RR)were used as metamodels to construct inversion models for LCC,leaf Anth,and NBI values of kiwifruit at different growth stages based on the Stacking integrated framework.Combined with drone RGB images,inversion mapping is carried out to achieve accurately estimation of physiology and biochemical parameters of kiwifruit leaves,thereby providing support for acquisition of kiwifruit growth information and refined orchard management.The results had indicated that:(1)As the growth period progresses,the LCC of kiwifruit slowly increased,with the highest content at maturity;The Anth value and NBI value of kiwifruit leaves showed a trend of first increasing and then decreasing,reaching their maximum value during the fruit expansion period.Different forms of mathematical transformation can highlight the spectral characteristics of kiwifruit leaves,the first derivative spectrum highlights the"Red Edge"information,the reciprocal spectrum swaps the"Peak Valley"features of the original spectrum,the square root transformation improved the spectral reflectance value,and the logarithmic spectral reflectance is negative.(2)Four mathematical transformation of the original spectrum could improve the correlation between spectral reflectance and kiwifruit LCC,leaf Anth value,and NBI value,with the first derivative transformation having the best effect.The optimized spectral index constructed based on the first derivative spectral screening of any two band combination has better correlation with the three factors than the selected traditional vegetation index and trilateral parameters.The Green Red Vegetation Index(GRVI)and Greenness Normalized Vegetation Index(GNDVI)had good correlations with the LCC,Anth value,and NBI value of kiwifruit leaves;The yellow edge amplitude,yellow edge area,red yellow normalization index,and red yellow ratio index were sensitive trilateral parameters of the three.(3)The Recursive Feature Elimination(RFE)algorithm could select the optimal feature subset and achieve data dimensionality reduction.Compared with the modeling full feature collection,the feature reduction ratios of the best modeling subsets for kiwifruit LCC,leaf Anth value,and NBI value were above 54%,65%,and 65%,respectively.(4)Stacking integrated model had higher estimation accuracy than single model.The optimal inversion periods for LCC and NBI values of kiwifruit leaves were both in the mature stage,and the optimal inversion models were RFE-Stacking RR models,with R~2 of 0.821 and 0.836,and RMSE of 0.292 and 3.095,respectively;The Anth value of kiwifruit leaves was suitable for inversion during the young fruit stage using the RFE-Stacking LR model,with estimated R~2 and RMSE of 0.938 and 0.010,respectively.(5)The correlation between the LCC,leaf Anth value,and NBI value of mature kiwifruit and the RGB color index of 33 drones were analyzed.The color index with the strongest correlation with the three was R-B,and the order of correlation from high to low was LCC(-0.587),leaf NBI value(-0.547),and leaf Anth value(0.544).The kiwifruit LCC estimation model constructed based on mature kiwifruit drone RGB images had better performance than the leaf Anth value and NBI value estimation models.The best estimation models for the three were RFE-Stacking RR models,and the modeling accuracy from high to low is:kiwifruit LCC(R~2=0.781,RMSE=0.332),kiwifruit leaf NBI value(R~2=0.673,RMSE=0.023),kiwifruit leaf Anth value(R~2=0.724,RMSE=4.277),this model was used to invert and map the LCC,Anth,and NBI values of kiwifruit based on RGB images,the results obtained could basically reflect the distribution of kiwifruit in the study area and provide a reference basis for kiwifruit field management.