Integrating Remotely Sensed Imagery and Existing Multiscale Field Data to Derive Rangeland Indicators: Application of Bayesian Additive Regression Trees

Abstract

Remotely sensed imagery at multiple spatial scales is used increasingly in conjunction with field data to estimate rangeland indicators (e.g., vegetation cover) and meet the growing need for landscape-scale monitoring and assessment of rangelands. Remote sensing studies that produce rangeland indicators often require intensive and costly field-data collection efforts to produce accurate model predictions. Existing monitoring data, such as those collected by the Bureau of Land Management’s Assessment, Inventory, and Monitoring (AIM) program, are potentially useful sources of field data in remote sensing modeling studies. Given their data-hungry nature, common regression tree−based modeling approaches may be inadequate for reliably predicting rangeland indicators with the smaller sample sizes of AIM data than typically used for remote sensing studies. Current literature suggests that Bayesian models, such as Bayesian additive regression trees (BART), may provide a suitable alternative to traditional regression tree−based modeling approaches to overcome the sample size limitation of the AIM data. In this study, we used 182 AIM field plots together with both high (RapidEye) and moderate (Landsat OLI) spatial resolution satellite imagery to predict bare ground and bare soil, total foliar, herbaceous, woody, and shrub cover indicators on rangelands in a 14 625-km2 area of northeastern California. We demonstrate that a BART model performed similarly to other regression tree approaches when field data and high spatial resolution imagery predictions were combined to predict indicator values using the medium spatial resolution Landsat image. The BART models also provided spatially explicit uncertainty estimates, which allow land managers to more carefully evaluate indicator predictions and to identify areas where future field data collection might be most useful. This study demonstrates that existing field data and freely available, remotely sensed imagery can be integrated to produce spatially explicit and continuous surface estimates of rangeland indicators across entire landscapes.