Research On The Nutrient Diagnosis And Prediction Of Flower Bud Differentiation Ability Based On The Hyperspectral Imaging Data

The hyperspectral images of Hamlin Sweet Orange (C. sinensis (L.)) leaves and flowers were collected respectively in flower bud differentiation stage (2013) and full-blossom stage (2014). Hyperspectral features of the citrus leaves and flowers were explored, and the nutrition diagnosis and flower number prediction technologies based on the hyperspectral data were developed. The main results were as follows:1. Nutrition diagnosis for citrus tree based on leaves’hyperspectral data(1) The hyperspectral images of spring and summer shoots in flower bud differentiation stage, and maturity spring shoots from the last year in full-blossom stage were acquired. The reflectance spectra (450-1000 nm) indicated that the reflectance spectrum fluctuation range reduced as increasing of leaves’ maturity along with the its’ stable metabolism.(2) Total of five spectra preprocessing methods (such as S-G smoothing, multiplicative scatter correction (MSC), Standard normal variate transform (SNV),1st Der, and 2nd Der) were employed to reduce the noise and scattering of raw spectra. The partial least square (PLS) models were developed to predict the contents of N, P, K and total soluble sugar (TC). The estimation results showed that the best preprocessing methods for the spring leaves and summer leaves in flower bud differentiation stage, and maturity spring leaves from the last spring season in full-blossom stage, were S-G smoothing,2nd Der, and S-G smoothing respectively.(3) The contents of N, P, K and TC in the leaves were accurately predicted by the leaves’ hyperspectral data. The optimal samples for N diagnosis (Rp=0.735), P diagnosis (Rp=0.733) and K diagnosis (Rp=0.728) were all the spring leaves in flower bud differentiation stage. The optimal sample for the TC diagnosis (Rp=0.688) was the spring shoot from the last spring season in flowering stage.2. Nutrition diagnosis for citrus tree based on flowers’ hyperspectral data(1) The extraction technique of region of interest(ROI) of flower hyperspectral images was explored. The ROI can be segmented founded on the image at 670.66 nm using dual-threshold method (0.5-0.6). The 692.2-753.3 nm feature wave range interval (FWRI) and 440.07-481.07nm, and 692.2-753.3 nm feature interval wave bands combination (FIWBC) was screened for for the N prediction of Hamlin flowers. The study found that the synergy interval partial least square (siPLS) model based on FIWBC got a higher prediction accuracy (Rp=0.762) by using 108 variables, reduced 85.8% of global spectrum.(2) The 573.757～633.437 nm FWRI, and 573.757-633.437nm, 693.74-752.653nm FIWBC for P prediction in Hamlin flower were chosen. The synergy interval partial least square (siPLS) model based on FIWBC was developed for More accurate prediction (Rp=0.885) by using only 152 variables.(3) The FIWB 593.313～643.711 nm for K prediction in Hamlin flower was better. The interval partial least square (iPLS) model based on the FWRI was developed for the better prediction (Rp=0.916) by using 64 variables.3. C/N ratio of leaf and flower number prediction using hyperspectral data(1) The feature wavelengths (507.117,507.886,523.294,527.928,534.114, 543.41nm) for C/N ratio of Hamlin leaf prediction were selected using interval-successive projections algorithm (SPA), and PLS model was achieved with Rp=0.914.(2) The flower number in canopy could be estimated using the spectra of maturity spring leaves from the last spring season in full-blossom stage. The feature wavelengths (544.186,552.726,567.516,572.196,575.319,582.352,588.613, 593.313nm) were selected for the flower number prediction by using interval-successive projections algorithm (SPA), and PLS model was achieved for the better flower number prediction (Rp=0.893).