Land cover classification : refining training requirements for support vector machine classification using remotely sensed data

Abstract

First, an SVM analysis was evaluated against a series of classifiers with particular regard to the effect of training set size on classification accuracy.  For each classification, accuracy was positively related with training set size.  In general, the most accurate classifications were derived from the SVM approach, and with the largest training set the SVM classification were more accurate (93.75%) than that derived from the discriminant analysis (90.00%), decision tree (90.31%) and artificial neural networks (92.18%).  The SVM classifier used about 50 per cent of the training data as support vectors.

If the regions likely to furnish support vectors could be identified prior to the classification, it may be possible to intelligently select useful training samples.  This was explored for the classification of agricultural crops in Feltwell area of U.K.  The support vectors of one of the crops, wheat, were mainly derived from peat soils.  Thus the ability to target useful training samples, in this case, based on soil type may allow accurate classification from small training sets in case the analysis is repeated in future.

 The training data requirements may be reduced if there is a prior knowledge or some ancillary information that can be used to identify/locate training sites to regions from which the most informative training samples, the support vectors can be derived.  This allows an intelligent training acquisition scheme to be devised in advance of training acquisition process and should include the variables affecting the spectral response of the classes.  This was demonstrated for agricultural classes in south western part of Punjab state of India.  Considering all the growth stages of the crops and background properties (water and soil) of the training sites provided appropriate support vectors central to the establishment of SVM classifier.  The scheme was successful in its intent to capture support vectors directly from field as 70% of the training samples collected were used by SVM as support vectors as compared to 47.7% for conventional training scheme.  The intelligent scheme of training data acquisition was cheaper by 26.09 per cent over the conventional scheme of training data acquisition because of reduced training set size. 

The training data requirements can also be reduced when the concern is to map accurately only one class from the many land cover classes available in the study area.  In such instances training data should be limited to the class of interest and classes facing the class of interest in feature space.  This was demonstrated here in accurately mapping cotton crop.

More information

Identifiers

Catalogue record

Export record

ASCII CitationAtomBibTeXData Cite XMLDublin CoreDublin CoreEP3 XMLEndNoteHTML CitationHTML CitationHTML ListJSONMETSMODSMPEG-21 DIDLOpenURL ContextObjectOpenURL ContextObject in SpanRDF+N-TriplesRDF+N3RDF+XMLRIOXX2 XMLReferReference ManagerSimple Metadata