At northern latitudes where the effects of arctic amplification drive warming at twice the rates seen in lower latitudes, climate patterns drive complex feedback loops in wetland ecosystems. The enormous reserves of ancient carbon locked into northern latitude soils are at risk of release under a rapidly changing climate. A key driver in soil carbon dynamics may be the shift from non-vascular to vascular plants; rates of carbon sequestration have been shown to be related to the distribution of vegetation type. We developed remote-sensing methods to scale from ground to airborne, and up to satellite resolution, to enable mapping and monitoring the changing patterns of northern wetland vegetation. Alaska’s Kenai Peninsula Wildlife Refuge was used as a test of our approach, because it comprises a complexity of sub-arctic ecosystems for which exist rich and underexploited hyperspectral and forest inventory datasets. On the ground, fieldwork included vegetation sampling and measurement of edaphic (soil) parameters and spectral reflectance. These data were compared to hyperspectral imagery from NASA’s G-LiHT campaigns. Ordination and spatial statistics such as Partial Least Squares Regression (PLSR) were used to model spatial predictions of vegetation composition as a function of spectral reflectance.
Results/Conclusions
We found strong statistical associations between wetland class and pH (p-value < .001), electrical conductivity (p-value < .001) and volumetric water content (VWC) (p-value < .01). In the Sphagnum-dominated wetlands, strong correlations between pH and VWC support the idea that Sphagnum drives hydrology and acidity in a negative feedback loop that further lowers pH. Early results from ground collected spectral data suggest we can discriminate between plant functional types using spectral profiles. PLSR was used on the handheld data to model the relationship between the reflectance profiles of the plant functional type and the ordination scores of the sampled plots. The regression resulted in an R2 for the observed values of 0.86, and cross validation R2 of 0.67 for predicted data. We scaled the methods up to the airborne sensor data from NASA’s G-LiHT imagery and applied the results of PLSR model pixelwise to a subset of a G-LiHT tile. The preliminary results suggest that we can we can use these methods to predict the distribution of plant functional types with relatively good accuracy; validation with fieldwork will be done in the coming field season.