Estimating aboveground carbon in a catchment of the Siberian forest tundra: Combining satellite imagery and field inventory Hans Fuchs, Paul Magdon , Christoph Kleinn, Heiner Flessa Remote Sensing of Environment 113 (2009) 518–531 Seo Hwan-seok

INDEX KEYWORD INTRODUCTION MATERIALS AND METHOD RESULTS DISCUSSION CONCLUSIONS

Multiple linear regression Keyword ASTER/Quickbird Carbon estimation Feature selection Forest inventory Forest tundra Global change k-NN regionalization Multiple linear regression

Introduction Forest ecosystems are an important part of the global carbon cycle because they store a large part of the total terrestrial organic carbon and exchange CO2 with the atmosphere. The transitional region of the forest tundra is expected to be sensitive to even small changes in environmental variables and thus it can offer early insight into potential changes in vegetation and aboveground carbon stocks driven by climate change. Two primary methods for the assessing of phytomass : Field measurements Indirect assessments using remote sensing - Linear Regression Model - k-Nearest Neighbor

Introduction Three Objectives Estimated and regionalize aboveground dry biomass and carbon stocks Apply the k-NN-technique for estimation and regionalization of AGC as well as to quantify and model ecosystem fluxes and budgets. Compare the utility of data obtained from medium and high resolution satellite sensors for AGC estimation.

Materials and methods Study area Area : 45 ha Location : Central Siberia near the town of Igarka Latitude : 67°48′ Longitude : 86°43′ Species : Betula pendula Picea obovata Variety of landscape units and vegetation associations in a distinctive spatial pattern

Materials and methods Structure of the carbon mapping process

Materials and methods Data Field data - date : 2006. 08 - Circle plot size : Radius = 7m - Random sampling (8point from each stratum) Remote sensed data - Quickbird : date = 2006.07.16, GSD = 0.6m(panchromatic), 2.4m(multispectral) view angle = 6.2°, sun elevation/azimuth = 44°/179.4° - ASTER : date = 2001.09.10, GSD = 15m(VNIR), 30m(SWIR), 90m(TIR) cloud cover = 2%, view angle = 2.7°, sun elevation/azimuth = 27.5°/180.7°

Materials and methods Image preprocessing Resampling Image transformations - Quickbird : 34 - ASTER : 27 Masking - Thermokarst ponds were excluded from the target set

Materials and methods Linear regression modeling - Stepwise regression k-Nearest Neighbor prediction Evaluation - Leave-one-out-cross-validation - Bias, RMSE, Relative Bias, Relative RMSE, Coefficient of variation Target sample Training sample K = 3 K = 7 (1) 참조표본점의 선정 목표 표본점과 참조 표본점의 밝기값의 유사성 (2) 참조표본점별 거리 가중치 목표표본점과 참조표본점 밝기값이 유사할수록 큰 가중치 부여 (3) 목표표본점의 추정치 참조표본점의 실제 측정치와 각 참조 표본점별 산출된 가중치를 이용해 산출

Results Feature selection - To get an overview of the relevance of the variables we explored the Pearson correlation coefficients r between AGC and the analyzed satellite image variables

Results Linear models - linear models for Quickbird and ASTER based on the selected variables The goodness of fit of the selected models was examined with analysis of scatter diagrams and residual graphs Show increasing residuals with increasing carbon values Comparison of predicted versus observed values reveals a clearly curved band pattern for ASTER and a slightly curved pattern for Quickbird raises the question of the suitability of the selected linear model

Results k-NN approach Minimum RMSEr for Quickbird 1 at k=10 Smaller k values preserve range and variance but result in greater uncertainty of predictions as the RMSEr is higher Because we wanted to maintain precision while conserving most of the variability of predicted carbon we used k=5 neighbors as a compromise

Results k-NN approach The residual graphs for the selected subsets 1 illustrate that the variance of the residuals increases with increasing values of predicted total carbon The predicted versus observed values give an indication of a clear but weak relationship for both ASTER and Quickbird

Results Comparison between linear regression and k-NN

Results Comparison between linear regression and k-NN Estimated AGC - Field sampling : 15.30 t/ha - Quickbird : k-NN : 14.70 t/ha Linear : 15.64 t/ha - ASTER : k-NN : 14.94 t/ha Linear : 14.41 t/ha

Discussion The carbon predictions from k-NN and linear regression are well in the range of values reported by other studies - Vederova et al. (2002) reported biomass values of 25.7 t/ha and 36.9 t/ha - Alexeyev & Birdsey (1998) reported an AGC of 15 t/ha - Nilsson et al. (2000) estimated an AGC of 23.9 t/ha - IIASA Forestry program (IASA FOR, 2007) published maps of the aboveground living phytomass density (dry matter) with a range of 8–23 t/ha at the location of the study area Little Grawijka Creek Catchment. The performance of image transforms such as vegetation indices, texture and tasseled cap transforms was compared against the original image bands

Discussion The stratified sampling produced sufficient reference data to be used for prediction using the satellite imagery, because the spectral range of the target set was covered well and we were able to compare estimates for strata zones with the regionalization results of k-NN prediction and linear regressions The panchromatic band of high resolution satellite images often yields better results for forest biomass and for prediction of stand characteristics than multispectral bands; this is in agreement with our results Analyzing the AGCmaps provided by the different methods one can see that k-NN has the effect of edge preserving and enhancement, while linear regression estimation provides a generalization in the spatial domain The computational effort for both methods can be reduced by using feature selection algorithms such as linear stepwise selection, but remains high when k-NN is applied for high resolution imagery like Quickbird

Conclusions The k-NN algorithm delivers reliable overall results as compared to estimates from stratified sampling Feature selection results in a considerable decrease of estimation errors Comparison of linear regression and k-NN prediction showed advantages for the regression approach as bias and RMSE were smaller and correlation coefficients of all linear models were higher than those of the corresponding k-NN imputations Linear regression estimation performs worse than k-NN when it comes to extreme observations : a general solution to the problem of negative prediction values for biomass is still to be identified in particular when dealing with small sample sizes

Thank you