Hyperspctral Diagnosis Of Nitrogen Content Level In Rubber Tree Under Naturally State Of Growth

Natural rubber is an important strategic resource in China,and its output has an important impact on China’s national economic level and international status.Rational fertilization based on the nitrogen content of rubber trees plays an important role in improving rubber production and improving rubber properties.Hyperspectral analysis technology has gradually become an important technical means for plant nitrogen detection due to the close relationship between its spectrum and phytochemical composition.Compared with traditional empirical judgment and chemical analysis,hyperspectral analysis technology has become an inevitable requirement for precision fertilization of natural rubber forests due to its high efficiency,non-destructive and accurate characteristics.This paper based on the in-depth analysis of hyperspectral image and its nitrogen content,discussed the correlation within the hyperspectral image band and the pixel respectively,proposed the Correlation Based Successive Projection Algorithm,established the deep learning based nitrogen diagnostic model with hyperspectral.The main innovations are as follows:(1)Correlation Based Successive Projection AlgorithmBy analyzing the principle of the original SPA,it is found that there may still be high collinearity between the non-adjacent bands extracted by the algorithm.Correlation analysis found that performing SPA in a region with consistent correlation under projection upper limit can extract a combination of wavelengths with lower collinearity but higher information retention.Result shows that the subset selected by CB-SPA has fewer feature variables than the original SPA,but they have the same distribution trend,meanwhile the correlation coefficient gray-scale map and VIF value show that variables selected by CB-SPA is significantly lower in collinearity than the original SPA.What’s more,the CB-SPA modeling time is 1/12 of the original SPA,and its prediction model has better performance both on the calibration and prediction set.(2)Deep Learning Based Nitrogen Diagnostic Model with HyperspectralBased on problem that traditional plant nutrient detection model with hyperspectral only model with the meanspectral,and not fully utilized the spatial and spectral information of hyperspectral,this paper proposed the Deep Learning Based Nitrogen Diagnostic Model with Hyperspectral—SAE-FNN.The Stacked Autoencoder is layerwised pre-trained with leaf pixel spectrum and fine-tuned with error back propagation after attached with FNN.The network prediction effects under different network layers,different input scales and data processing conditions are compared and analyzed.The results show that the best predict model is established on region-averaged hyperspectral with a 2800 pixel regions selected and a 4 layer network.The root mean square error is 0.0547.(3)Adaptive Region Averaging AlgorithmDealing with the high correlation between the spectra of adjacent pixels on the hyperspectral image of rubber tree leaves,an adaptive region averaging algorithm is proposed to mean the spectra of a region once all spectra of the region is determined to be correlated significantly.Therefore,lower correlated hyperspectral image could obtained with not much loss of spatial and spectral information.After treated by the adaptive region averaging algorithm,ore lower correlation subset of pixel spectra could be utilized by the SAE-FNN network,so that improving the predict accuracy and stability of the deep network.