CLASSIFICATION OF ENVIRONMENTAL DATA WITH KERNEL BASED ALGORITHMS (PREPRINT IBRAE-2002-09)

Preprint IBRAE-2002-09

Pozdnukhov A., Timonin V., Kanevski M., Savelieva E., Chernov S.

Soil type classification is an important problem from different points of view. Vertical migration of radio-nuclides in soils can be mentioned as an example. The process of migration depends on a number of different properties corresponding both to radio-nuclides and soils. All soil properties are strongly connected with a soil type. Official soil type maps are not good enough to be used for migration problems. Real soil type is more variable value, than it is usually presented in official maps. Soil type mapping can be improved by using additional information obtained during radio-nuclide concentration measurement.

In this work the classification problem is solved by machine learning methods such as probabilistic neural networks PNN (supervised learning algorithm) and Support Vector Machines SVM. The advantages of both methods are general non-linear modelling that avoids the direct modelling of spatial correlation structure. The methods are compared with the nearest neighbour method, the simplest approach to spatial classification.

Bibliographical reference 

Pozdnukhov A., Timonin V., Kanevski M., Savelieva E., Chernov S. CLASSIFICATION OF ENVIRONMENTAL DATA WITH KERNEL BASED ALGORITHMS. Preprint IBRAE-2002-09. Moscow: Nuclear Safety Institute RAS, 2002. 23 p. — Refs.: 10 items.