Study On Extraction Of Sensitive Bands Of Conifer Species Death Based On Hyperspectral Data

Masson’s pine and Chinese fir are the most important coniferous tree species in China,which have important economic,ecological and social value.In recent years,due to natural disasters and large area outbreak of pine wood nematode disease,the safety of timber stands has been greatly affected.Therefore,it is very important to monitor and study the diseases,insect pests and disasters of conifer species.In this paper,Chinese fir and Masson pine were selected as research objects to collect the hyperspectral data of the tree death process,and the spectral data were transformed by first-order differential,second-order differential,logarithm,normalization,standard normal conversion and other algorithms.Moreover,the sensitive bands of conifer species when diseases and insect pests occurred were extracted by combining T-test and particle swarm optimization algorithm.The extracted bands were classified by spectral Angle classification algorithm,support vector machine classification algorithm and Bayesian classification algorithm.After comparing the extraction results of 5 data transformation algorithms combined with the two bands,the following conclusions were drawn:(1)The spectral changes of the death process of Cunninghamia lanceolata and Masson pine were similar,and the earliest bands that changed were mainly from 550nm to 710nm.(2)The T-test can effectively identify and extract the sensitive bands of the death process of Chinese fir and Masson pine.The speed of dimension reduction of PSOSVM is relatively fast,but the classification results are not ideal.Moreover,the bands extracted based on the current stress data only have a good classification accuracy for the current stress data,while the classification accuracy is low for the data of other stress periods of the same trees.It does not apply to the entire process of detecting dead trees.(3)Due to the low signal quality of the original spectrum and the large amount of redundant information and noise,the number of bands extracted by combining T-test is much less than that extracted by other data transformation algorithms.Therefore,it is necessary to preprocess the original spectral data to improve its separability and accuracy.Different algorithms have different transformation effects on the original spectrum.NR algorithm can highlight the small differences in the spectrum,making the band extraction more sensitive at different stress stages.Log algorithm has a good effect on band extraction in different stress stages and can identify important features in the spectrum.SNV algorithm is sensitive to band extraction in near infrared region,but weak in visible region.FD1 and FD2 algorithms are relatively insensitive to spectral noise and chaotic information,and introduce a lot of noise while amplifing spectral differences.In practical applications,FD1 and FD2 algorithms need to be optimized to improve their accuracy and stability in plant spectral feature extraction.Support vector machine classification is the best,followed by spectral Angle classification,Bayes classification algorithm performance is not stable.(4)By observing the original spectral changes of the death process of Chinese fir and Masson’s pine,it can be seen that the spectral changes of the death process of Chinese fir and Masson’s pine are similar,and it can be inferred that the spectral changes of the death process of different conifer species are very similar.The sensitive bands discussed in this paper are as follows:520nm-809nm,929nm-1053nm,1474nm1609nm,1746nm-1782nm,1996nm-2200nm.