This paper proposes a modified subspace classification method (mscm) that integrates an extended multiple similarity method with the averaged learning subspace method to achieve improved land-cover classification performance. Furthermore, we developed an automatic parameter selection and optimization technique for MSCM to avoid time-consuming, laborious manual parameter tuning. Only three parameters need to be set, and their optimal values are easily determined by the automatic procedure. We carried out experiments with data of two multispectral images: Landsat TM data for a semi-arid area in the Horqin sandy land, China, and ASTER data for the Kasumigaura Lake region, Japan, a high humidity, warm temperature zone. Accuracy assessment of the mscm results in comparison with those of the support vector machine (svm) and maximum likelihood classification (mlc) methods showed that the MSCM yielded better classification results. Therefore, the proposed MSCM shows promise as a tool for land-cover classification.
ASJC Scopus subject areas
- Computers in Earth Sciences