Land cover classification using satellite -sensed imagery and its texture values: An accuracy assessment based on the Florida Land Use and Cover Classification System

Computer-automated classification of remotely sensed imagery from satellites has proven useful for applications in land cover evaluation and land use planning. However, the lack of accuracy in land cover recognition hinders the usual spectral differentiation method. One technique to enhance accuracy of land cover recognition is the use of texture images, defined as a set of local statistics or other local properties of an image which are constant, slowly varying, or approximately periodic.;In this study, two test sites near Tampa, Florida, were used to test whether the incorporation of the texture measures within the computer automated classification techniques increases classification accuracy. LANDSAT-TM and SPOT-Panchromatic images were merged to optimize spatial and spectral resolution. Reference land covers were taken from existing classifications according to (a) the Florida Land Use and Cover Classification System (FLUCCS) and (b) the Florida Game and Fresh Water Fish Commission (FGFWFC)---land cover classification system. Error matrices and the Kappa coefficients of agreement were used to describe classification accuracy. Twenty-seven different texture measures (three non-directional and six directional, using 3x3 window size for each of the spectral bands) were computed and used along with the spectral bands as variables for supervised, unsupervised, and hybrid methods of land cover classifications. In subsequent analyses highly correlated bands were merged using principal component transformation, land cover classes with close proximity (as measured by the Jefferies-Matushita distances) were combined, and a Bayesian maximum-likelihood classifier was also added to enhance classification accuracies. Finally, a new mixed classification method was developed and applied, allowing separate band combinations for each of the land cover classes. The classification accuracies obtained by the Gaussian maximum-likelihood method using only the spectral bands were taken as reference, and accuracies of other methods were compared with them.;Use of all the spectral and texture bands together produced lower land cover classification accuracies as compared to the reference values. The Bayesian method resulted in better classification. The methods with highest accuracy differed between the sites. The hybrid technique of classification yielded poor results, while the mixed combination classification method improved land cover classification results. In most cases, texture variables added only marginal information towards differentiating land cover classes.;The results show a modest increase in accuracies of the classifications. The increased accuracies indicate the marginal usefulness of the texture measures under the present constraints. However, the study indicates enough merit in texture measures to warrant further research of its potential.