Monitoring inter-annual woody-cover change in savanna ecosystems: applying random forests to “scalable” predictor variables

Authors: Rebecca Powell*, University of Denver, Sydney M. Firmin, University of Denver
Topics: Remote Sensing, Geographic Information Science and Systems, Arid Regions
Keywords: savannas, percent tree cover, random forests, vegetation change, Landsat, MODIS
Session Type: Paper
Presentation File: No File Uploaded

The defining characteristic of savanna ecosystems is the coexistence of grasses and woody vegetation (i.e., trees and shrubs). The spatial distribution of woody cover has implications for wildlife habitat and rural livelihoods, and directly affects ecosystem function in terms of energy exchange, nutrient cycling, and carbon storage. Yet, fundamental questions remain about woody cover dynamics in savanna systems, including whether changes in woody cover occur incrementally or abruptly in space and time. Our goal is to develop a robust methodology to predict percent woody cover in savanna ecosystems - particularly in areas that have very low woody cover (<30%) - using freely available, moderate spatial resolution imagery. First, we first develop a percent woody cover base layer for Serengeti National Park, Tanzania, using a random forests (RF) regression model applied to 30-m Landsat 8 surface reflectance and DEM-derived variables. Next, we test the “scalability” of the RF regression model as applied to 500-m MODIS surface reflectance and ancillary data. We examine the efficacy of model predictions based on the number and timing of input image dates, as well as the impact of per-pixel fire return intervals. Accuracy of each product is assessed using reference data from two sources: a database of manually interpreted Google Earth imagery (90-m samples), and classification of very high spatial resolution multispectral imagery. Our results represent a substantial improvement over current global models of percent tree cover, in part because our methods are optimized for sparsely vegetated environments with low-stature woody cover.

To access contact information login