Remote Sensing Classification Of Stellera Chamaejasme In Degraded Alpine Meadows

Alpine meadows are an important part of grasslands,making important contributions to ecological balance,water conservation,and grassland production.As a major grassland type on the Qinghai-Tibet Plateau,alpine meadows have been affected by global climate change and human activities in recent years,with different degrees of degradation trends.The poisonous weeds represented by Stellera Chamaejasme L.are spreading rapidly in the area and seriously threaten the meadow ecosystem with Cyperaceae Juss,and Kobresia Willd vegetation as the dominant species.Therefore,the use of remote sensing technology to obtain information on the spatial distribution of Stellera Chamaejasme L.communities in alpine degraded meadows on a large scale,rapidly and accurately is essential for making decisions on the control of alpine degraded meadows.In this paper,a typical wolfsbane occurrence area in the middle part of Qilian County,Qinghai was selected as the study area,and the Kruskal-Wallis test & Dunn post hoc test,Classification and Regression Trees(CART)and Jeffries-Matusita(J-M)distance algorithm extracts Stellera Chamaejasme L.white flowers,flower buds and green plant leaf spectra,as well as Stellera Chamaejasme L.samples with different cover degrees and green grass samples,and then uses the J-M distance to adjust the community classification results based on two-way indicator species analysis(TWINSPAN)to obtain the Stellera Chamaejasme L.remote sensing identification scheme.Different machine learning methods are combined to construct classifiers to classify Planet multispectral images and to evaluate the classification accuracy.The main findings of this paper are as follows:(1)Combined with Kruskal-Wallis & Dunn post-hoc test,CART and hierarchical analysis of J-M spectral distances extracted feature identification bands with J-M distances >1.9 at both the plant flower\leaf spectral scale and the sample spectral scale,with an overall average dimensionality reduction efficiency of 91.14%.Among them,38 bands were identified between Stellera Chamaejasme L.white flowers-green leaves and 4 bands between Stellera Chamaejasme L.flower buds-green leaves.There were 9 bands between high-coverage Stellera Chamaejasme L.community samples and grass community samples,9bands between medium-coverage Stellera Chamaejasme L.samples and grass samples,and 3bands between low-coverage Stellera Chamaejasme L.samples and grass samples.Combining the results of leaf and sample scale analysis,the characteristic identification bands between Stellera Chamaejasme L.communities and other grassland communities were mainly located in 392～540nm、567～576nm、581～739nm、1127～1134nm、1148～1209nm、1243～1373nm、1383nm、1409～1460nm、1595nm、1600nm、1616～1647nm、1657nm、1662～1718nm 、 1729～1734nm 、 1746～1749nm 、 1754 nm 、 1879 nm 、 1885～1950nm 、1956nm、1964～1967nm、1975～1992nm、2001～2029nm、2042～2051nm。(2)The TWINSPAN quantitative class classification divided the 93 samples into 6communities,I: Kobresia pratensis + Kobresia alpine + Cassia lanceolata + Potentilla bifida+ Gentiana macrophylla;II: S.chamaejasme + Kentucky bluegrass + Potentilla pratensis + C.lanceolata + Gentiana;III: S.Stellera + K.bluegrass + Gentiana + alfalfa + Cassia lanceolata;IV: K.bluegrass + S.Stellera +bluegrass + alfalfa + Elymus przewalskii.V: E.przewalskii +alfalfa + dandelion + P.anserina + plantain,VI: Plantago przewalskii + Plantago asiatica +Potentilla anserina + Caragana.The samples with spectral conflicts in TWINSPAN classification results were adjusted based on J-M distance,and were divided into 4communities,category II、III、IV and grassland communities(I+V+VI).And linking spectral features with ecological community distribution to construct a spectrally distinguishable plant community classification scheme.(3)The classifiers were constructed using random forest,support vector machine,Knearest neighbor method,and maximum likelihood method to classify Planet images for wolfsbane identification,and the classifier model accuracies of random forest,support vector machine,K-nearest neighbor method,and maximum likelihood method were 88%,75%,75%,75%,respectively,and total accuracy of classification were 75.47%,66.04%,72.17%,68.88%,classification accuracy of Stellera Chamaejasme L.were 74.24%,65.15%,71.97%,65.15%.And The applicability of the four classification methods for remote sensing identification of wolfsbane in alpine degraded meadows was ranked from largest to smallest:random forest > K nearest neighbor > maximum likelihood method > support vector machine.From the classification results,Stellera Chamaejasme L.has become one of the dominant species in the area,with patchy distribution on the sunny slopes of mountain ranges,roads and both sides of rivers,and large random scattered distribution in low-lying and flat areas,which can provide a reliable basis for remote sensing identification of Stellera Chamaejasme L.in alpine degraded meadows.