Quantifying tree cover in an African savanna using a multi-scale remote sensing approach

Background/Question/Methods

Savanna is widespread globally and is the dominant African biome.  Tree cover is a key attribute of ecosystem structure, a dominant driver of ecosystem function and biodiversity, and plays an important role in socio-ecological systems.  Understanding process and pattern in savanna tree cover dynamics is critical to the health of these ecosystems and requires appropriate tools for monitoring tree cover change across space.  Remote sensing provides a useful tool for analyzing landscape patterns at multiple spatial and temporal scales, but identifying suitable algorithms for quantifying tree cover variation in savanna ecosystems has proved challenging.   We present a multi-scalar technique for quantifying percent tree cover using IKONOS and MODIS imagery combined with field data from the Serengeti ecosystem in Tanzania.   We applied an unsupervised classification to the 1 m IKONOS panchromatic band (February 15, 2010) and a low pass convolution, filtered image to isolate a tree-cover class in our field site. We upscaled this tree cover map to the coarser resolution of the MODIS13Q1 250-m product, and used a 2-year (IKONOS date plus/minus 1 year) time series of MODIS 16-day composites to identify suitable metrics for quantifying tree cover at a coarser resolution.  We used three sets of metrics: longitudinal summary statistics (e.g., mean and standard deviation), Fourier harmonics, and NDVI during dry-season greenup. Analysis was done using four remotely sensed variables (NDVI, RED, MIR, and NIR MODIS bands), as well topographic variables.  We further validated our model using Google Earth imagery extracted from a random population of MODIS pixels across the Serengeti.

Results/Conclusions

Tree cover ranged between 0 and 46% at 250-m resolution.  The IKONOS classification was strongly correlated with field-measured values of tree cover (R² = 88%).  The MODIS calibration set (N = 2530) showed a reasonably strong correlation (R² = 71%) with IKONOS tree cover (validation set: N = 1295, R² = 71%).  The strongest overall predictor of tree cover at the MODIS scale was NDVI at the time of late-dry season greenup, which corresponds to woody leaf flush prior to the onset of the rainy season.  This suggests that using time series of NDVI data in association with phonological differences among functional groups is far more reliable than non-NDVI based metrics for quantifying tree cover in this savanna system.