Influence of spatial aggregation on prediction accuracy of green vegetation using boosted regression trees

Abstract

Data aggregation is a necessity when working with big data. Data reduction steps without loss of information are a scientific and computational challenge but are critical to enable effective data processing and information delineation in data-rich studies. We investigated the effect of four spatial aggregation schemes on Landsat imagery on prediction accuracy of green photosynthetic vegetation (PV) based on fractional cover (FCover). To reduce data volume we created an evenly spaced grid, overlaid that on the PV band and delineated the arithmetic mean of PV fractions contained within each grid cell. The aggregated fractions and the corresponding geographic grid cell coordinates were then used for boosted regression tree prediction models. Model goodness of fit was evaluated by the Root Mean Squared Error (RMSE). Two spatial resolutions (3000 m and 6000 m) offer good prediction accuracy whereas others show either too much unexplained variability model prediction results or the aggregation resolution smoothed out local PV in heterogeneous land. We further demonstrate the suitability of our aggregation scheme, offering an increased processing time without losing significant topographic information. These findings support the feasibility of using geographic coordinates in the prediction of PV and yield satisfying accuracy in our study area.