Post-processing of Pixel and Object-Based Land Cover Classifications of Very High Spatial Resolution Images

Abstract

The state of the art is plenty of classification methods. Pixel-based methods include the most traditional ones. Although these achieved high accuracy when classifying remote sensing images, some limits emerged with the advent of very high-resolution images that enhanced the spectral heterogeneity within a class.

Therefore, in the last decade, new classification methods capable of overcoming these limits have undergone considerable development. Within this research, we compared the performances of an Object-based and a Pixel-Based classification method, the Random Forests (RF) and the Object-Based Image Analysis (OBIA), respectively. Their ability to quantify the extension and the perimeter of the elements of each class was evaluated through some performance indices. Algorithm parameters were calibrated on a subset, then, applied on the whole image. Since these algorithms perform accurately in quantifying the elements areas, but worse if we consider the perimeters length, hence, the aim of this research was to setup some post-processing techniques to improve, in particular, this latter performance.

Algorithms were applied on peculiar classes of an area comprising the Isole dello Stagnone di Marsala oriented natural reserve, in north-western corner of Si-cily, salt pans and agricultural settlements. The area was covered by a World View-2 multispectral image consisting of eight spectral bands spanning from visible to near-infrared wavelengths and characterized by a spatial resolution of two meters. Both classification algorithms did not quantify accurately object perimeters; especially RF. Post-processing algorithm improved the estimates, which however remained more accurate for OBIA than for RF.