A spatial ensemble approach for broad-area mapping of land surface properties

Abstract Understanding rapid global change requires land cover maps with broad spatial extent, but also fine spatial and temporal resolution. Developing such maps presents a unique challenge, as variability in relationships between spectral characteristics (i.e., predictors) and a response variable is likely to increase with the size of the region across which a model is built and applied. Although most mapping approaches apply the same predictor-response relationships globally across the entire modeling region, learned relationships from one local area may be invalid for another when predicting across broad extents. Here, we adapted a spatial ensemble approach borrowed from species distribution modeling to land cover mapping, and evaluated whether the approach could faithfully represent spatial variation in relationships between land cover and spectral data. The spatiotemporal exploratory model (STEM) uses an ensemble of regression trees defined within spatially overlapping support sets, producing a broad-extent map that reflects variability at the spatial scale of each constituent support set. As test cases for reference maps, we used 30-m-resolution forest canopy and impervious surface cover layers from the 2001 U.S. National Land Cover Database (NLCD) for the states of Washington, Oregon, and California. When testing strategies for support set size and sampling intensity, we found that predictor-response relationships were strongest when individual components of the spatial ensemble were small and when sampling intensity was high. Compared to aspatial bagged decision tree and random forest models, we found that the STEM approach successfully captured variation in our source maps, both globally and at scales smaller than the modeling region. Leveraging the spatial structure of a STEM, we also mapped per-pixel spatial variation in prediction confidence and the importance of different predictor variables. After testing appropriate spatial ensemble and sampling strategies, we extended the predictor-response relationships gleaned from the 2001 source maps into a yearly time series based on temporally-smoothed spectral data from the LandTrendr algorithm. The end products were yearly forest canopy and impervious surface cover time series representing 1990–2012. Formal evaluation showed that our temporally extended maps also closely resembled NLCD maps from 2011. The aim of this research was to cultivate the implicit relationships between spectral data and a given map, not improve them, but as the need for time series maps produced at both broad extents and fine resolutions increases, our results demonstrate that an ensemble of locally defined estimators is potentially more appropriate than conventional ensemble models for land cover mapping across broad extents.