Abstract: 【Objective】To study a technology for estimating the number of apple flowers which is based on the hyperspectral image information of a canopy at full-bloom stage, in order to lay a foundation for the establishment of the technologies used for the management and the productivity prediction of the flowers and fruits of a plant. 【Method】 The 5-year-old Malus pumila ‘Mitch Gala’ trees with M₉ clonal rootstocks in the shape of high spindles were studied. The visible and near infrared hyperspectral images of the canopy at full-bloom stage were collected, and the number of the flowers of the trees was selected and then counted manually. Finally, comparatively analyze the effects of the Partial Least Squares (PLS) models based on the original reflectance spectra (OS) and the spectra pretreated by five kinds of methods including savitzky–golay smoothing(SG), standardization of normal variables (SNV), Normalize, first derivation(l^st Der), second derivation(2^nd Der), the PLS, the back-propagation neural network (BPNN) and the least squares support vector machines (LS-SVM) based on characteristic wavelengths obtained by x-loading weight (x-LW) on the accuracy of the real-time estimation of the amount of flowers per unit area per tree. 【Result】 Both the number of flowers per tree and the number of flowers per unit area per tree have high correlation coefficients, which means using the number of flowers per unit area of the canopy as a substitute for the total number of flowers per tree to predict the number of flowers of all the trees is feasible. The number of flowers per unit area per tree had a very significant positive correlation with the reflectivity of the trees’ canopy in the ultraviolet and visible wavelength (308-700 nm), but the correlation between the two was not significant in the near-infrared wavelength (750-1 000 nm). Based on full wavelength, the PLS model based on spectra pretreated by Normalize predicts the number of flowers per unit area per tree most accurately, whose determination coefficients of the calibration sets (R_c²) and of the prediction sets (R_p²) were 0.794 and 0.804, the root-mean-square errors of the calibration sets (RMSEC) and of the prediction sets (RMSEP) were 0.084, 0.062, and whose relative error (RE%) of prediction sets was 3.940. Based on characteristic wavelengths, the stability of BPNN model is bad，the LS-SVM model has better modeling results, with the R_c² and the R_p² being 0.826 and 0.804, RMSEC and the RMSEP being 0.077 and 0.064, the RE% of prediction sets is 12.160. 【Conclusion】 The PLS prediction model based on Normalize pretreatment has the best prediction results of the number of flowers of apple trees in the shape of high spindles per unit area per tree’s canopy, at the same time, after obtaining the spectral data by hyperspectral imager, analysis and processing to extract feature information through simplification, which can provide a basis for the application of multispectral remote sensing data.

Key words: apple tree, flowers per unit area, hyper-spectra, PLS, x-LW