How up-scaling of remote-sensing images affects land-cover classification by comparison with multiscale satellite images

Land-cover classification provides the crucial data component for related Earth-science research. Currently, although multiscale remote-sensing images are the main source of data for classifying land-cover, the response of multi-resolution images to land-cover classification remains highly uncertain. In addition, because of the scarcity of time-synchronous multiscale satellite images of certain regions, up-scaling algorithms are generally used to generate and apply multiscale images. However, when using resolution-resampling images, it remains uncertain to what extent spectral loss or information distortion is responsible for the underlying differences in the accuracy of land-cover classification of various landscapes. To clarify this situation, we study the Hetian basin of Changting County in Fujian Province, south-east China by using quasi-synchronous multiple-resolution satellite observations (seven spatial resolution levels: 1 m, 2 m, 4 m, 8 m, 16 m, 30 m, and 50 m) to investigate possible correlations between spatial resolution and the land-cover classification. The classification is obtained by applying a support vector machine spectral classifier to random recordings made in 1875 sample plots. We also explore the effect of using lower-resolution images by comparing the classification results obtained by using several common up-scaling algorithms, such as nearest neighbour (NN), bilinear (BI), cubic convolution (CC), and pixel aggregation (PA). The results indicate that classification accuracy is significantly influenced by the spatial resolution of images (p < 0.05), with the accuracy increasing as the spatial resolution goes from 1 m to 4 m, then decreasing as the spatial resolution decreases beyond 4 m. In addition, for a resolution of 1 m to 30 m, almost all the up-scaled images provide a classification accuracy that differs from that obtained by using the native remote-sensing images of each resolution (p < 0.05), and the difference increases as the spatial-resolution ratio or up-scaling amplification factor increases. According to an analysis of the spatial scale of images using, e.g., multiband spectral reflectance and vegetation index, the up-scaling algorithms are less sensitive to spatial resolution and represent poorly the actual image characteristics. This result is attributed to the strong dependence of the spectral information in up-scaled images on the original images, which leads to discrepancies with respect to actual observations at the given scale. These results indicate that the effects of resolution cannot be ignored and that resampling data may not be adequate for multi-spatial-scale classification compared with the native satellite images. It is thus urgent to obtain an effective up-scaling algorithm that sharply reduces the problems caused by spatial heterogeneity.