Comparative Studies on Similarity Distances for Remote Sensing Image Classification

Scene classification is one of the most important tasks in the remote sensing field. In general, remotely sensed data comprises targets of different nature with many detailed classes. Therefore, the classification of patches in a satellite scene is a challenging issue. To address the problem, the preferred alternative is to transform to polar coordinates and analyze angular distances. Prior works have so far considered angular distances between points, while ignoring that the target class is not a point, but a distribution. In this paper, we take advantage of this critical fact by using a point-to-probability distribution measure rather than an $\ell_{n}$ norm. In this paper, two similarity measures (Euclidean and Mahalanobis) in two different feature space are experimentally investigated through some remote sensing datasets.

1. Introduction

Remote sensing (RS) imagery is a valuable data source for Earth Observation to measure different physical properties on earth. Compared to natural images, RS imagery consists of various wavelengths and provides more information. As a result, understanding massive and complex RS multispectral images is critical and more challenging. There are some reasons why satellite image processing is more challenging: (1) big intraclass diversity, (2) high interclass similarity, (3) large variance of object/scene scales, and (4) coexistence of multiple ground objects [1]. Recently, many RS tasks have been carried out by exploiting optical multispectral images from Sentinel-2, which perform very well when cloud coverage is minimal. When the weather is cloudy, this technique is not viable since cloud removal itself is a challenging task [2]