The Study Of Vegetation Classification Based On Hyperspectral Images In Zhalong Wetland

Wetland is one of the most important ecological system with great ecological andeconomic value. But with the rapid economic development of human beings and theirrational use of water resources, the area of our country’s natural wetland is reducedyear by year, the original landscape pattern are damaged and the result directly affectthe distribution of vegetation in the wetland. In recent20years, people have are-understanding to the value of the wetland. So all kinds of comprehensive researchand protection of vegetation were paid more and more attention. The traditional wetlandvegetation classification based on multispectral image generally, so it’s hard todistinguish different surface features when they have similar spectral features and thusmake a confusion of Classification Results. And hyperspectral images have hundreds ofnarrow band, and it makes a possibility that distinguish the similar spectral features ofground targets, which traditional remote sensing cannot. So it promotes the vegetationclassification accuracy of the research area.Based on the summary of domestic and abroad wetland vegetation classificationresearch, this paper take Zhalong wetland as study area and launch the jobs as renderedbelow: Samples of spectral curve of the wetland vegetation were collected in the wildas the main basis for the classification process including weeds kind of meadow, leymuschinensis, cultivated land, paddy field, reed, leaf of cattail, deep-water vegetation;Geometric correction, atmospheric correction and some other preprocessing weretaken to the remote sensing image applied; A fusion of hyperspectral remote sensingimage and CCD image were made in order to improve the resolution ofthe hyperspectral remote sensing image. The fusion algorithm include HSI transform,principal component analysis, Brovey transform, and Gram-Schmidt transformation.According to the purpose of fusion, from improving the spatial resolution spectralproperties improve clarity and keep improving information four aspects and fiveevaluation index, including mean standard deviation information, entropy averagegradient and related coefficient to evaluate the several fusion method of quality; On thebasis of fusion image， supervised classification and non-supervised classificationmethod of ISODATA algorithm， K-means algorithm， maximum likelihood，minimum distance method and spectral Angle mapping method was used. In the studyof vegetation distribution classified, according to the general classification accuracy andKappa coefficient comparing the different classification accuracy of these classificationmethods. There are three result in this article as followed on the basis of the methodabove:(1) The main vegetation types in the study area (weeds kind of meadow, leymuschinensis, cultivated land, paddy field, reed, leaf of cattail, deep-water vegetation) havedifferences spectral features. Weeds kind of meadow has the highest reflectance and deep-water vegetation has the lowest，so they can be used as the main basis for theclassification based on the hyperspectral remote sensing image.(2) The results of different sources of remote sensing image fusion can providemore comprehensive image classification and more accurate more reliable referenceinformation. It not only enhance the spatial resolution but also keep the advantage oforiginal hyperspectral image, this makes the classification results can be used for futuredecision planning with more meaningful reference.(3) The classification results show that in all the classification method, themaximum likelihood method has the lowest overall classification accuracy and Kappacoefficient, the reason may be due to the sample of spectral features a normaldistribution has certain deviation; In ISODATA algorithm, paddy, planting and reedscan papyrus grow tall where three types were divided into the meadow wrong number ismore that could be due to the three planting types of spectral features and meadowsspectral features (especially when soil background when water content are biggermeadow) caused by relatively close; In K-means algorithm,29.5%of the submergedvegetation types were divided into reeds and rushes, the three planting types are inwater as a background, its reason is probably because the water spectral features cause aconfused result with these three plant; Minimum distance classification method morepaddy fields and rushes was wrong into reed, but this classification method of theoverall accuracy is higher, at80.71%(Kappa coefficient=0.775), in other conditions ofvegetation classification is also worthy to use; Spectral Angle mapping method of theoverall classification accuracy is89.86%(Kappa coefficient=0.8817) was significantlyhigher than the other classification method of classification accuracy.