Evaluations of landcover classification characteristics for Landsat TM and SPOT HRV data were performed. Three kinds of classification experiments using the most popular supervised maximum likelihood classifier were conducted. The first one is an assessment of effects of higher spatial resolutions, in which images of several ground resolutions (20m-75m) were classified. The purpose of the second experiment is to clarify the usefulness of higher spectral resolution of TM data. In this experiment, several band combinations were tested. The last one is the evaluation of spatial feature characteristics. Four kinds of texture features were extracted from co-occurrence matrices, and the effectiveness of those features for a landuse classification were evaluated. As a result, the following conclusions were obtained. The addition of new spectral bands, i.e. band 1, 5 and 7 in TM has introduced a little increase in the landcover classification accuracy. The increase of spatial resolutions does not necessarily provide higher classification accuracies. This fact indicates that some kinds of spatial informations should be utilized to obtain higher accuracies. However, a simple addition of texture features to spectral features could not increase the classification accuracy.
It has been shown from recent studies that conventional pixelwise supervised land-use/cover classification methods can not achieve sufficient classification accuracies for high ground resolution image data such as SPOT HRV and Landsat TM. In this paper, a new landcover/use classification algorithm in order to increase classification accuracies for high resolution images is presented. This algorithm is based upon a two stage recognition model of landcover/use. In this model, each concept of landcover categories is defiend by the component ratio of landcover elements surrounding the corresponding pixel. The classification procedure can be divided into 3 steps ; the first step is the land cover elements recognition of each pixel using a pixelwise classifier; second step is the calculation of the component ratio of each element within local image region ; third step is the final decision for landcover categories using a minimum distance classifier. The renults of experiments showed that this algorithm achieved about 7% improvements of classification accuracies.