Discriminating Stellera chamaejasme in alpine grasslands using UAV hyperspectral imagery and multivariate analysis

Abstract

Stellera chamaejasme is highly invasive and causes a significant threat to alpine grassland on the Qinghai-Tibet Plateau. It is important to determine its distribution pattern in order to stem the rapid invasion of this toxic weed. In the present study, strategies for mapping S. chamaejasme based on UAV Resonon hyperspectral imaging were assessed in combination with dimension reduction, clustering and ordination analysis, and spectral separability measurement. Field spectrometry analysis showed that the hierarchical procedure of Kruskal-Wallis test and Dunn’s post hoc test, CART and JM distance can efficiently select the minimum optimal wavelengths for S. chamaejasme discrimination and considerably reduce the dimensions of hyperspectral imagery. DCA and RDA ordination revealed that S. chamaejasme identification at the species level is difficult even using hyperspectral feature bands. The reaggregating of TWINSPAN ecological communities based on the criterion of JM distance > 1.9 can result in spectrally delineated S. chamaejasme communities and the co-existing species community. Applying three algorithms of MLC, RF, and SVM, the performance of S. chamaejasme classification based on JM-TWINSPAN schemes was significantly improved compared to that based on TWINSPAN schemes. Of these, RF and JM-TWINSPAN achieved the best classification result (OA = 91.00%, Kappa = 0.83, AD = 14.0%, and QD = 1.2%). The result indicates that the promising accuracy can be achieved in accurate mapping of S. chamaejasme by a multivariate approach, which combines ecological analysis and separability measurement with image classification.