Fine Vegetation Classification And Fractional Vegetation Cover Retrieval Using Time Series Remote Sensing Data

Functioning as a vital part in ecosystem,vegetation plays a crucial role in the carbon sink around the globe.And therefore,vegetation has long been a highlyconcerned target in the research of global water-carbon-nitrogen cycling.Remote sensing monitoring,on the one hand,can acquire information of vegetation types,forest area as well as coverage on a relatively large temporal and spatial scales.And on the other hand,it is of great significance in the monitoring of the changes.The problems existing in current research of vegetation can be summarized as follow: Firstly,there is the lack of remote sensing data's continuity.Besides,there are obstacles in the fine classification of vegetation type.Thirdly,there are difficulties in obtaining vegetation field survey data.What's more,the retrieval methods of fractional vegetation cover is rough.To all the issues above,this paper proposes a method for fine vegetation classification and fractional vegetation cover retrieval.The method is based on time series remote sensing data by synthetically applying phenomenological parameter fitting,multi-year time series data synthesis,JSEG multi-scale segmentation,machine learning and other research methods.And the main conclusions are as follow:(1)The fine classification of vegetation based on MODIS time-series key phenological parametersThe phenological phenomenon in northern part of Hulubuir demonstrates distinct characters.The growing season of the forest has an earliest start(145-160 DOY)and it ends also relatively early(250-275 DOY).As for the growing season of the grassland,the start(160-170 DOY)is slightly lagged behind that of forest while the overall length is the same as that of the forest.The growing season of the crop can be described as short and intensive,which starts late(170-195 DOY)and ends early(225-285 DOY).An overall accuracy of 81% has been achieved by applying the decision tree method to classify the key parameters of the phenological phase.(2)The fine classification of vegetation based on multi-year Landsat time-series synthetic dataFrom January to March,the EVI values of evergreen coniferous forests,evergreen broadleaf forests and shrubs in Guangdong province present various change trend.And the same goes with the EVI values of the planted forests,shrubs and evergreen broadleaf forests from July to August.The EVI value of farmland,however,is distinct from that of other vegetation types.An overall accuracy of 80.03% has been achieved by applying the random forest method to classify multi-year Landsat time-series synthetic data.The importance ranking of the random forest's feature parameter indicates that in Guangdong Province,the phenological change characteristics of vegetation from winter to spring are the important determinants of the difference of vegetation types.The summer phenology information and DEM information are also important factors to distinguish different vegetation types and besides,it is main factor to distinguish vegetation area from non-vegetation area as well.(3)The vegetation cover extraction based on multi-scale JSEG method and unmanned aerial vehicle imagesThe results of the research by using multi-scale JSEG method segmenting UAV images show that different segmentation scales have their own advantages.Scale-1 is conducive to the segmentation and extraction of a single crown,and scale-4 is the most suitable scale for discriminating light and shaded area.The overall accuracy of forest crown extraction in shaded area is slightly worse than the overall accuracy of light area;the overall accuracy of forest crown extraction obtained from densely distributed sampling points is slightly lower than the overall accuracy of sparsely distributed sampling points.In the case of JSEG segmentation using scale-4,the optimal total forest crown area extraction accuracy was achieved using a 9 × 9 opening filter for all the four combinations of planting density and light conditions.For all sample points,the overall accuracy of crown area extraction is 90%.(4)Fractional vegetation cover retrieval based on multi-year synthetic Landsat time-series dataThe results of fractional vegetation cover retrieval show that the EVI data in July is the most suitable data for stepwise regression retrieval of fractional vegetation cover.The EVI data in May and September also play an important role in the stepwise regression retrieval of fractional vegetation cover.In May-August,the EVI data obtained higher importance scores in the random forest regression retrieval of fractional vegetation cover,which reflected the importance of summer data for the retrieval of fractional vegetation cover.The correlation coefficients of evergreen broadleaf forest and planted forests fractional vegetation cover retrieval using random forest regression were 0.52 and 0.5,respectively,and the correlation coefficient of shrub fractional vegetation cover retrieval was 0.27.The correlation coefficient of evergreen coniferous forest fractional vegetation cover retrieval using support vector machine regression was the best and the correlation coefficient was 0.59.From the view of vegetation's distribution,the proportion of evergreen broadleaf forests shows a tendency of increasing with elevation while the proportion of planted forests,on the contrary,performs a decreasing tendency.The proportion of evergreen coniferous forests shows no obvious relation with the elevation.The fractional vegetation cover in the study area in Guangdong shows a great spatial difference as follow: no vegetation covered parts and low fractional vegetation covered parts exist primarily in the plain area;high fractional vegetation covered parts are mainly distributed in the higher altitude area.The vegetation types with the highest average fractional vegetation cover in the study area of Guangdong were planted forests,followed by evergreen broadleaf forests and evergreen coniferous forests,while the mean fractional vegetation cover of shrub was the lowest.The study found that the forest coverage in Guangdong study area was 52.46% from 2014 to 2016,which is consistent with the eighth national forest inventory in Guangdong(51.26%).Through the above research,this study,first of all,effectively compensates for the lack of time series information of remote sensing data for a single year.And secondly,the study has achieved an overall improvement in the precision of fine classification of vegetation.What's more,the study has significantly improved the precision of the extraction of fractional vegetation cover in shadowed areas under complex surface conditions.Last but not least,the study has more accurately and efficiently obtained the result of fractional vegetation cover of forest.To sum up,this study provides more theoretical basis and methodological reference for the fine classification of vegetation and retrieval of fractional vegetation cover.