Validating classification accuracy of low spatial resolution data by using high spatial resolution data

Abstract

This is a common practice to use imagery classification method to assess vegetation changes. The accuracy of classification of remote sensing data affects objective assessment of changes, but high accurate classification is difficult to obtain through the use of low spatial resolution and visual interpretation techniques. Once classified, these land use and land cover maps need to be validated to assess how accurate they are using field in-situ survey data. Validation with in-situ data, however, is quite labor intensive and sometimes economically impossible. Therefore, a simple and easy way of estimating classification accuracy of a low spatial resolution map is a still difficult problem. In this paper, we use high spatial resolution data to validate the classification maps derived from low resolution data. Using NOAA AVHRR images and Landsat-TM over the same area acquired in the same year, we developed a method to validate vegetation classification maps from the NOAA image with classification maps derived from high resolution TM images. Once scaled to the same spatial resolution, TM-image derived maps were compared with those from NOAA AVHRR classification data. The results suggested that over two separate areas that we compared; the accuracies of classification images were 0.804 and 0.793, respectively, suggesting that the classification accuracy is nearly as high as 80% from the NOAA AVHRR images. The method presented here is very useful in assessing accuracies of classification maps derived from coarse resolution images such as those from AVHRR data.