FOREST TYPE AND ABOVE GROUND BIOMASS ESTIMATION BASED ON SENTINEL-2A ANDWORLDVIEW-2 DATA EVALUATION OF PREDICTOR AND DATA SUITABILITY

摘要

The present study analyses the two earth observation sensorsrnregarding their capability of modelling forest abovernground biomass and forest density. Our research is carriedrnout at two different demonstration sites. The first isrnlocated in south-western Germany (region Karlsruhe) andrnthe second is located in southern China in Jiangle Countyrn(Province Fujian). A set of spectral and spatial predictorsrnare computed from both, Sentinel-2A and WorldView-2rndata. Window sizes in the range of 3*3 pixels to 21*21rnpixels are computed in order to cover the full range ofrnthe canopy sizes of mature forest stands. Textural predictorsrnof first and second order (grey-level-co-occurrencernmatrix) are calculated and are further used within a featurernselection procedure. Additionally common spectralrnpredictors from WorldView-2 and Sentinel-2A data suchrnas all relevant spectral bands and NDVI are integratedrnin the analyses. To examine the most important predictors,rna predictor selection algorithm is applied to the data,rnwhereas the entire predictor set of more than 1000 predictorsrnis used to find most important ones. Out of the originalrnset only the most important predictors are then furtherrnanalysed. Predictor selection is done with the Borutarnpackage in R (Kursa and Rudnicki (2010)), whereas regressionrnis computed with random forest. Prior the classificationrnand regression a tuning of parameters is done byrna repetitive model selection (100 runs), based on the .632rnbootstrapping. Both are implemented in the caret R packagern(Kuhn et al. (2016)). To account for the variability inrnthe data set 100 independent runs are performed. Withinrneach run 80 percent of the data is used for training and thern20 percent are used for an independent validation. Withrnthe subset of original predictors mapping of above groundrnbiomass is performed.