Spatial Classification for Land Cover Mapping

Zhe Jiang, [email protected] Abstract: My research explores novel computational techniques to map the physical cover (e.g., forests) of the earth’s surface from satellite images. Processing these images is labor-intensive and a significant burden on scientists. Existing methods ignore spatial information and assume that pixels are statistically independent. Consequently, they produce erroneous maps that require prohibitive manual refinements. In contrast, I propose a novel spatial-contextaware approach in computer science. My approach reduces the burden of map refinement by capturing each pixel’s spatial neighborhood properties. The approach is also scalable on large datasets. Preliminary results have confirmed that the proposed approach works in real-world applications.