关键词: 沉香, 全氮, 多光谱, Elastic Net, 图像处理

Abstract:

Objective: This study applied computer vision technology to process the multispectral images of young Aquilaria sinensis canopy, and build a total nitrogen estimation model, to realize the nondestructive estimation of the nitrogen nutrition status of Aquilaria sinensis and explore the new image processing methods and new estimation model for Aquilaria sinensis nutritional monitoring. Method: Based on the canopy images of Aquilaria sinensis seedlings obtained by the multispectral camera, the improved exponential semi-soft threshold function was used to wavelet denoise the multispectral image, and the Fourier-Merlin transform (FMT) registration and genetic algorithm-back propagation neural network (GA-BPNN) algorithm were combined to achieve accurate segmentation of the multispectral image. Then, the spectral texture features of the segmented images were extracted as independent variables to establish the Elastic Net (EN) total nitrogen estimation models improved by artificial bee colony algorithm (ABC), beetle antennae search algorithm (BAS), and hybrid gray wolf optimization algorithm (HGWO). Finally, the model accuracy was verified and evaluated by using the leave-one-out cross-validation method to determine the optimal model, which was compared with the traditional partial least squares regression (PLSR), ridge regression (RR) and multiple stepwise regression (SR) models. Result: 1) The improved wavelet exponential semi-soft thresholding method could effectively remove strong noise in near-infrared images, and was superior to wavelet soft thresholding, hard thresholding and traditional semi-soft thresholding methods. 2) The segmentation effect of GA-BPNN was better than that of Otsu threshold, maximum entropy and minimum cross entropy segmentation, and the segmentation effect combined with FMT registration was significantly better than that of direct segmentation of images in various bands. 3) As for the EN models, based on the verification result of the single spectral feature, single texture feature and comprehensive feature models, the optimization effect of the three optimization algorithms was ranked as ABC>HGWO>BAS, and the estimation accuracy of the three feature models was ranked as comprehensive feature > single spectral feature > single texture feature. For the ABC-optimized EN models (ABC-EN), the determination coefficient (R²) of the comprehensive feature model was 0.829 4, which was 10% and 54% higher than that of the single spectrum and texture feature models respectively, and the mean square error (MSE) was 0.169 5, which was 30% and 63% lower than that of the single spectrum and texture feature models respectively. 4) The comparison result of estimation accuracy between ABC-EN and other traditional models were: ABC-EN > RR > PLSR > SR, and the R² of RR, PLSR and SR model were 0.748 3, 0.651 2, and 0.577 9, which were lower than the ABC-EN model we proposed. Conclusion: The wavelet exponential semi soft threshold denoising method can effectively remove the strong noise in the multispectral image of young Aquilaria sinensis, and the segmentation algorithm combined with FMT and GA-BPNN has achieved good segmentation result. The ABC-EN model built by fusion features of spectrum and texture is the best estimation model of total nitrogen in the canopy of aloe vera seedlings. Therefore, the method proposed in this paper has certain practical significance for monitoring the nitrogen nutrition status of Aquilaria sinensis, and is beneficial to the realization of precise operation in the process of Aquilaria sinensis cultivation and management.

Key words: Aquilaria sinensis, total nitrogen, multispectral, Elastic Net(EN), image processing