Modifying geographically weighted regression for estimating aboveground biomass in tropical rainforests by multispectral remote sensing data

The present study uses a local regression approach for estimation of aboveground biomass (AGB) in a tropical rainforest area with highly diverse terrain conditions from remote sensing-based multi-spectral vegetation indices (VI). By incorporating altitudinal effects into the spatial weighting matrices of the common geographically weighted regression (GWR), an extended GWR model, geographically and altitudinal weighted regression (GAWR), has been developed to deal with both spatial (horizontal) and altitudinal (vertical) non-stationarity in the data set. Unlike the common GWR model, the presented GAWR approach captures both horizontal and altitudinal drifts in the relationships between aboveground biomass and remote sensing data. In order to test its improved performance, the GAWR method was compared with the traditional GWR technique and global ordinary least squares regression (OLS) in a region of mountainous tropical rainforest in Sulawesi, Indonesia. The relationships between AGB and VIs were found to be significantly spatially variable. The results showed that there were substantial benefits in capturing both horizontal and vertical non-stationarity simultaneously. The GAWR method significantly improved AGB prediction in all simulations relative to both the traditional GWR and OLS methods, as indicated by accuracy and precision statistics. From the results of empirical tests, it seems proper to say that for this data set, the GAWR model is better than the traditional GWR model.