Spatial non-stationarity and scale-dependency of prediction accuracy in the remote estimation of LAI over a tropical rainforest in Sulawesi, Indonesia

The technique of Geographically Weighted Regression (GWR) was used for estimation of Leaf Area Index (LAI) from remote sensing-based multi-spectral vegetation indices (VI) such as Normalized Difference Vegetation Index (NDVI), the mid-infrared corrected Normalized Difference Vegetation Index (NDVIc), Simple Ratio (SR), Soil-Adjusted Vegetation Index (SAVI) and Reduced Simple Ratio (RSR) in a region of equatorial rainforest in Central Sulavesi, Indonesia. The linear regressions between NDVI, NDVIc, SR, SAVI and RSR as explanatory variables and ground measurements of LAI at 166 plots as a dependent variable were produced using common modelling approach — Ordinary Least Squares (OLS) regression fitted to all data points, as well as GWR. Accuracy and precision statistics indicate that the GWR method made significantly better predictions of LAI in all simulations than OLS did. The relationships between LAI and the explanatory variables were found to be significantly spatially variable and scale-dependent. GWR has the potential to reveal local patterns in the spatial distribution of parameter estimates, it demonstrated sensitivity of the modelʹs accuracy and performance to scale variation. The GWR approach enables finding the most appropriate scale for data analysis. This scale was different for each VI. The results suggest that spatial non-stationarity and scale-dependency in the relationship between LAI and remote sensing data has important implications for estimations of LAI based on empirical transfer functions.