Estimation And Spatial Scale Transformation Of Forest Above-ground Biomass Based On Landsat 8 OLI Images

Forest above-ground biomass(AGB)is one of the most important indicators for assessing forest ecosystem health,carbon sink potential and the effectiveness of sustainable forest management measures.However,due to the strong heterogeneity of forest structures,the characteristics of forest ecosystems at different scales within a large geographical area have obvious differences,and the local estimation results cannot accurately reflect the biomass distribution characteristics of the whole forest ecosystem.In light of this,a new framework that integrates multiscale geographic weighted regression(MGWR)and Co-kriging interpolation algorithm to perform the spatial downscaling operation of forest AGB based on 168 permanent sample plots’observations in partnership with the Landsat 8 OLI imagery in partial areas of Lishui City,Zhejiang Province,was proposed to achieve the goal of not affording additional higher spatial resolution remote sensing images to generate finer forest AGB spatial details.At the same time,the 15 m forest AGB pattern generated by the optimal downscaling method was upscaled to 90 m by using five upscaling methods,followed by an assessment of the upscaling performances of the five methods.Specifically,three sets of predictor variable sets were first identified by using random forest importance ranking,multiple stepwise regression and Pearson correlation coefficient combined with variance inflation factor methods,and then the set of predictor variables with the highest mean spatial smoothness was taken as the optimal downscaling variable set by taking the scale of action(bandwidth)of each predictor variable on the forest AGB into account.Based on this,the ordinary least square model(OLS),random forest model(RF),geographically weighted regression(GWR),multiscale geographic weighted regression(MGWR),and extreme gradient boosting model(XGBoost)were used to downscale the forest AGB,and the downscaling performance of the five models was compared quantitatively to determine the optimal downscaling model.Finally,the residuals of the predicted forest AGB downscaling results were analyzed by spatial interpolation on the basis of spatial autocorrelation among AGB samples,and the structured components of the residuals were extracted to overlay onto the predicted downscaling results to further refine the accuracy of the forest AGB prediction framework.In addition,the nearest neighbor method,bilinear interpolation method,cubic convolution method,local average method,and dominant variance weight method were used to perform the upscaling operation of the predicted 15 m forest AGB,and four evaluation metrics including mean,standard deviation,information entropy,and mean gradient were used to assess the upscaling performance of different methods.The main results of this study were as follows:(1)Both the GWR model and the MGWR model showed that the average and median bandwidths of the variable set selected by random forest importance ranking were the largest among the three sets of variables,indicating that they had the highest spatial smoothness when predicting the AGB.And the downscaling model built later double confirmed that this variable set had the strongest power in explaining the observed AGB variance.In this set of variables,both elevation and aspect affected forest AGB to a small extent and they were local variables,with a strong spatial heterogeneity.Some of the textural features and spectral features had larger bandwidths,with con＿B4＿3,B357,B14,var＿B5＿3 and brightness having bandwidths exceeding150 km,indicating that these variables affected forest AGB over a larger range.Spatial pattern analysis of the regression coefficients of this variable set showed that the regression coefficients of elevation and aspect varied substantially over space,indicating that both variables were the main factors influencing the spatial differentiation of forest AGB.B14,B16,B357,elevation,sec＿B7＿5,var＿B5＿3,and mean＿B4＿5 all showed positive effects on forest AGB across the entire area.In order to obtain the 15 m feature variable set,four image fusion methods were applied.The PCA method retained most of the information of the original image and better demonstrated the detail variation and texture features of remote sensing images compared to other three fusion methods.(2)Among the five downscaling models constructed in this study,the ordinary least square model(OLS)showed the worst prediction performance(with a validation R~2=0.46),and two machine learning models including random forest model(RF)and extreme gradient boosting model(XGBoost)obtained better prediction results than the OLS model,and the XGBoost model had lower residual squared and coefficient of variation,as well as higher R~2.The MGWR model exhibited the best prediction performance in all metrics among the five downscaling methods.In the study area with complex topography,taking aspect as a covariate to interpolate the prediction residuals derived from the MGWR downscaling model,and the results showed that the prediction accuracy of the Co-kriging model(CK)was higher than that of the ordinary kriging model(OK).Overall,the final validated R~2 of the proposed MGWRD-CK model was at 0.62,which effectively improved the spatial distribution characteristics and texture details of the forest AGB mapping.(3)Five upscaling methods(nearest neighbor,bilinear interpolation,cubic convolution,local averaging,and dominant variance weighting)were used to upscale the MGWRD-CK predicted 15 m AGB downscaling results to 90 m and the amounts of corresponding information loss and retention were compared.The results showed that the upscaling results based on the nearest neighbor method lost part of the image element values,the conversion results of the local average method appeared more noisy,and the upscaling results of the dominant variance weight method had the best performance and could retain most of the information features of the original image.(4)By considering the spatial non-smoothness,the spatial scale conversion results of AGB at two scales(15 m and 90 m)were effectively obtained,and they can provide basic data for the monitoring and management of regional forest resources and the assessment of carbon storage capability.