Research Of Bamboo Forest Information Extraction Using Multi-Scale Remote Sensing Data

Bamboo forest is a kind of forest resource type widely distributed in tropical and subtropical regions.Its processing is convenient,of good comprehensive benefits.Meanwhile,bamboo has a good ability of carbon sequestration,the assessment for capacity of bamboo forest’s carbon dioxide absorption in the global carbon cycle is based on knowing the distribution area of bamboo forest.Compared with other forests,bamboo whips out the bamboo shoots,and grows clonally,has the function of self-expansion in area,so that the traditional method of time-consuming to confirm the area and distribution of a wide range of bamboo forest often cannot meet the needs of the bamboo forest monitoring.Remote sensing monitoring well makes up for the deficiency of the traditional methods in this aspect.In this study,the main research object was bamboo forest,and the multi-scale remote sensing technique(Landsat 8 OLI,MODIS and other auxiliary data)was used to establish multivariate linear models at provincial scale in the form of classifying layer by layer,which took moderate resolution imaging results for reference data.The seasonal parameters were joined in the models,and the random forest method was applied in processing of feature selection and bamboo information extraction.The results from random forest method were compared to the new model,then the main factors affected the accuracy of bamboo estimation were analyzed.The conclusions can be drawn as follows:(1)The vegetation index NDWI based on near infrared and short-wave infrared wavelengths was more effective than other vegetation indices in bamboo forest information extraction.Compared with the vegetation index or matched filtering only,the method based on combination of vegetation index NDWI and the matched filtering had better ability for bamboo forest information extraction.This method provided a new try of bamboo forest classification mapping quickly and accurately.(2)For the MODIS data,the near infrared(845-885 nm)band and short wave infrared 1(1560-1660 nm)band from MODA1 regular bands,LWSI,SATVI and NDSVI from vegetation indices,the seasonal parameters,“base value” and “end of season”,from EVI time series were effectual to the bamboo forest information extraction at provincial scale;in the NDVI and EVI time series,the amplitude of bamboo forest in the growing season was greater than the broad-leaved forest.Taken together,EVI time series was more sensitive in vegetation growth change than NDVI time series.(3)To the random forest method,the bamboo forest information extraction result based on all the41 variables using was better than other variables combination results,followed by the random forest information extraction result using vegetation index variables.The EVI seasonal parameters in the bamboo forest information extraction was far superior to the NDVI seasonal parameters.To the multiple regression model method,the model using LWSI,SATVI,NDSVI,“EVI_base value” and“EVI_end of season” was more accurate than the model using regular bands(b2,b4,b5 and b7).Taken together,although the total accuracy of the method using five few variables was lower than the random forest method using all 41 variables,with 1.30% difference,it can greatly shorten the time and improve the efficiency of extraction,and it reduced the requirement of computer hardware configuration.(4)The remote sensing process,spatial pattern and the mixed degree of pixels will have an effect on the accuracy of the proposed model.In spatial landscape pattern,both at class scale and landscape scale,the accuracy of the proposed model was correlative with the metrics described the aggregation extent at class or landscape scale.The internal mixed degree of pixels can also affect the proposed model to extract bamboo forest information.