Modified average local variance for pixel-level scale selection of multiband remote sensing images and its scale effect on image classification accuracy

Abstract The development of remote sensor technology makes it convenient to obtain multiscale satellite data sets, but selecting data with an appropriate scale has become a problem. We propose improvements based on modified average local variance (MALV) for selecting the optimal spatial resolution of multiband images. One improvement is computing the mean MALVs of all bands, and the other is computing the average MALV of the selected bands. We discuss the optimum index factor and principal component analysis (PCA) methods for band selection. Further image classification experiments with different spatial resolutions are employed to verify the proposed methods. The experimental results prove that the MALV method is suitable for images with simplex landscape type. When the spatial extent of the image data is large, the MALV of the subimage whose landscape type is similar to the dominating landscape of the whole image is significantly referential for selecting the optimal spatial resolution. MALV based on PCA is more effective for reflecting the scale effect of spatial resolution and thus is useful for selecting the optimal spatial resolution of a multiband image. The experimental results also prove that very high spatial resolution will lead to high heterogeneity within class, and thus it will lead to low separability and low classification accuracy. Furthermore, the MALV method provides a feasible approach for quantitative research of the modifiable area unit problem.