Using Support Vector Machines to automatically extract open water signatures from POLDER multi-angle data over Boreal regions

Abstract

Boreal wetland ecosystems cover an estimated 90 x 106 ha, about 36% of global wetlands, and are a major source of trace gases emissions to the atmosphere [1]. Four to 20 percent of the global emission of methane to the atmosphere comes from wetlands north of 40°N latitude [2]. Large uncertainties in emissions exist because of large spatial and temporal variation in the production and consumption of methane. Accurate knowledge of the areal extent of open water and inundated vegetation is critical to estimating magnitudes of trace gas emissions. Improvements in land cover mapping have been sought using physical-modeling approaches [3],[4], neural networks [5],[6], and active-microwave [7], examples that demonstrate the difficulties of separating open water, inundated vegetation and dry upland vegetation. Here we examine the feasibility of using a support vector machine to classify POLDER data representing open water, inundated vegetation and dry upland vegetation.