Research On The Fine Classification And Chlorophyll Content Estimation Of Citrus Fruit Trees Based On Hyperspectral Data.

Real-time,non-destructive monitoring of distribution and physiological parameter information of citrus fruit trees is an important basis for systematically studying the planting structure and growth monitoring of citrus orchards.Machine learning algorithms have become one of the important means for accurate identification and growth monitoring of economic crops.However,systematic research exploring the applicability of machine learning algorithms in fine classification and chlorophyll content inversion of citrus fruit trees has not been conducted.In this study,the Litang Mockat citrus experimental base in Guilin was selected as the research area.Unmanned aerial vehicle hyperspectral imaging data was used to select and classify citrus fruit tree features through SULOV combined with the extreme gradient boosting(XGBoost)algorithm.Based on this,the quantitative relationship between ground-measured spectrum and leaf chlorophyll content(LCC)was investigated.The raw spectrum was processed using the fractional-order derivative(FOD)and the continuous wavelet transform(CWT).By using vegetation indices and constructing new band combinations,feature bands suitable for estimating leaf chlorophyll content were identified.The effect of FOD and CWT spectral processing methods on the inversion of LCC at the leaf scale was analyzed.Finally,the integrated learning regression(ELR)model was used to estimate LCC after feature processing was performed.The results obtained are as follows:(1)The combination of SULOV and the XGBoost algorithm effectively improved the accuracy of fine classification of citrus fruit trees.The algorithm proposed in this study successfully extracted significant features that reflected different citrus varieties during feature selection.The XGBoost classification algorithm achieved higher accuracy in the fine classification of tree crops with small feature differences and different varieties,resulting in an overall classification accuracy of 92.14%.The fusion feature of the first-order differential inflection point band and the original band had a significant impact on the fine classification of citrus.(2)The continuous wavelet transforms(CWT)proved to be effective in improving the estimation accuracy of leaf chlorophyll content(LCC).CWT eliminated the influence of overlapping peaks in the original spectrum on LCC estimation and enhanced the correlation between the original spectrum and LCC.The maximum correlation coefficient value of 0.894was obtained at scale 5.Even in a high-dimensional feature dataset,using only 5 vegetation index features in the ELR model inversion resulted in good accuracy(R~2=0.863).Vegetation index features based on scale 5 demonstrated significant superiority compared to other scales.(3)The fractional-order derivative(FOD)technique refined the changing trend of the raw spectrum,mitigated the impact of baseline drift and overlapping peaks,and considerably improved the correlation between the raw spectrum and LCC.The estimation accuracy of the double-band vegetation index,constructed based on FOD-processed hyperspectral data,was significantly lower than the three-band vegetation index combination method that fused leaf water content-sensitive bands.However,the ELR model optimized by the Hyperopt algorithm achieved higher inversion accuracy(R~2=0.876)in the estimation of chlorophyll content using the TBI2 single feature variable.Combining the optimal vegetation index of CWT with the optimal band combination feature of FOD yielded the highest accuracy(R~2=0.891).(4)The ELR model exhibited higher accuracy compared to any single machine learning regression algorithm.The model successfully integrated the advantages of sub-models and effectively mitigated the impact of sub-models’shortcomings,thereby improving the accuracy and robustness of LCC estimation using the ELR model.