Discussion on possibility of the identification of karst vegetation communities based on OLI data

Desertified karst region is a focal area of vegetation recovery and ecological restoration in southwest China, and vegetation is an important and sensitive factor to reflect the changes of ecological environment in karst region. Recently, with the development and application of imaging spectrometer, remote sensing technology plays an important role in large-scale karst vegetation investigation. Remote sensing data have the advantage of macroscopic, real-time, dynamic and make vegetation investigation of large areas more convenient. Zhongliang Mountian was taken as the study area, which is an area that vegetation is recovering in karst rocky desert area of Chongqing. The image of Landsat 8 OLI data in August 19, 2013 was applied as the test data in the study. we tried to discuss the effect and feasibility of BP artificial neural network which is a kind of neural networks using multi-spectral images to classify and identify the karst vegetation, the method includes neural network based on the multi-spectral bands of remote sensing image, or accompanied by non-remote sensing information, including texture information, slope, aspect and elevation. Finally, the supervised classification based on maximum likelihood was applied to extract the vegetation classes which were compared with the classes extracted by BP neural network. The results showed that Cupressus funebris, Cunninghamia lanceolata (Lamb.) Hook and mixed broadleaf-conifer forest couldn't be identified well with neural network when only multi-spectral data were employed. With the application of other data, such as ratio band, texture, slope, aspect and elevation, the total accuracy was improved gradually, and the three vegetations were well identified. It was concluded that the effect of neural network classification(NNC) based on multi-spectral bands of OLI image and DEM is the best in which the total accuracy reached 87.42%, the Kappa coefficient 0.85. Compared the results with that of traditional supervised classification, the total accuracy of NNC has shown a growth of 5.57%. For mapping accuracy, the identification accuracy of Pinus massoniana, bamboo forest, broadleaved forest, shrubland and hassocks were relatively high in which the mapping accuracy reached 91.03%, 94.03%, 84.39%, 80.50% and 99.21% respectively. While one of Cupressus funebris, Cunninghamia lanceolata (Lamb.) Hook and mixed forest was low, which might be because the area of three vegetation communities in the study area was relatively small, which led to self learning effect is poor. As a result, the accuracy wasn't always ideal. In general, the OLI multi-spectral remote sensing image using BP neural network is feasible to identify the Karst vegetation communities, and provide a scientific method for vegetation mapping. At the same time, it should be pointed out that the difference of spectrum characteristic of vegetation community is small or the remote sensing image has mixed pixel, this will cause loss of some precision. Aiming at the deficiency, we can combine multi-sensor remote sensing data with non-spectral information to extract information efficiently.