Spatial extensions to kernel methods
Alexei Pozdnoukhov · 2010
Kernel methods became one of the mainstreams of machine learning research in recent decades. This success is due to their ability to provide robust non-linear models with good generalization abilities which are easy to train and interpret. Spatial heterogeneity, prior knowledge on spatial similarities, discontinuities, physical and administrative boundaries can be incorporated into kernel modeling frameworks but require special consideration. This paper describes a general framework for building spatial extensions to kernel methods via data-driven kernel transforms. It is illustrated numerically by constructing spatial extensions to kernels for environmental and remote sensing applications.