Predicting apple tree leaf nitrogen content based on hyperspectral and wavelet packet analysis

This research is aimed at exploring high accuracy method on detecting nitrogen content for apple leaves in different physiological phenological phases. The experiments were conducted during the periods of fruit-bearing, fruit-falling and fruit-maturing separately. 20 apple trees were selected randomly from different regions in an apple orchard located in Beijing suburb, China. Then a main branch of each target tree was selected and three representative parts (base part, middle part and top part) of every bough were marked. And then leaves samples were collected from each representative part of each target tree, and 60 leaves samples were obtained in each phenological period. The collected samples were carried to the laboratory quickly, and their visible and NIR spectral reflectance were measured using Shimadzu UV-2450 spectrograph and their nitrogen content were detected using Kjeldahl method. For data processing, firstly data cluster analysis was conducted among the spectral reflectance and nitrogen content based on individual tree, hence 20 new sample data were obtained accordingly. Then the spectrum of each tree was decomposed using wavelet packet technology. The results revealed that with the wavelet packet decomposition scale increasing, signal of spectrum low-frequency and de-noised high-frequency separated gradually. The low-frequency signal became smoother apparently, some peak-valleys reflecting the biological characteristics disappeared. For the de-noised high-frequency signal, it didn’t change significantly with decomposition scale deepened in the visible region, while the noise decreased in the near infrared region. And then principle component analysis was applied respectively to the original spectra, extracted low-frequency spectra and de-noised high-frequency spectra. Finally, linear regression models for predicting leaf nitrogen content were established based on the principle components extracted from the according spectra and NDVI (859nm, 364nm). The results indicated that: (1) in different psychological phonological phases, the total nitrogen content forecasting models built with different wavelet packet decomposition spectra had higher accuracy than that with NDVI since full spectra could reserve more valid information than the signals at two sensitive wavebands; (2) the models established using the principal components extracted from the de-noised high-frequency spectra had the highest accuracy in fruit-bearing and fruit-maturing period. While in physiological fruit-falling period, the model established by the principal components extracted from the low-frequency spectra was the best; (3) in fruit-bearing period, the highest accuracy regression model went to which established based on the principal components extracted from the high-frequency noise removed spectra after 5-layer decomposition. Its calibration R2 reached to 0.9502, RMSEC was 0.0978, and the validation R2 reached to 0.7285, RMSEP was 0.0885; (4) in fruit-falling period, the best regression model went to that established based on the principal components extracted from the low frequency spectra after 7-layer decomposition. Its calibration R2 reached to 0.9539, RMSEC was 0.0553, and the validation R2 reached to 0.9273, RMSEP was 0.087; (5) in fruit-maturing period, the best regression model was that established based on the principal components extracted from the high-frequency noise removed spectra after 3-layer decomposition. Its calibration R2 reached to 0.9577, RMSEC was 0.0576, and the validation R2 reached to 0.9013, RMSEP was 0.0791; (6) wavelet packet decomposition technique is an effective way to enhance the spectrum prediction ability of apple tree leaves nitrogen content, meanwhile in order to improve the predicting accuracy, wavelet packet decomposition level should be determined based on the spectral characteristics in different physiological phonological phases.