High-resolution Urban Image Classification Using Extended Features

Abstract

High-resolution image classification poses several challenges because the typical object size is much larger than the pixel resolution. Any given pixel (spectral features at that location) by itself is not a good indicator of the object it belongs to without looking at the broader spatial footprint. Therefore most modern machine learning approaches that are based on per-pixel spectral features are not very effective in high- resolution urban image classification. One way to overcome this problem is to extract features that exploit spatial contextual information. In this study, we evaluated several features in- cluding edge density, texture, and morphology. Several machine learning schemes were tested on the features extracted from a very high-resolution remote sensing image and results were presented.