Multi-dimension and multi-granularity segmentation of remote sensing image based on improved Otsu algorithm

Abstract

An increasing number of unknown islands, an important resource for human development, is identified based on segmentation of remote sensing image. Different from traditional digital image, remote sensing image has significant characteristics, such as multi-band, multi-source, and multi-granularity. Thus, the segmentation theory based on traditional digital image is not suitable for remote sensing image. Here, the segmentation algorithm (Otsu), which is a common method for traditional digital image, was improved in two aspects: (1) Based on PCA and band fusion, the Otsu algorithm with one-dimensional image was extended to multi-dimensional ones; (2) By optimizing the threshold value, the Otsu algorithm for single feature extraction was extended to multi-granularity extraction. Taking the island segmentation from remote sensing image as an example, the improved Otsu algorithm was compared with the traditional Otsu: 1) Through using PCA algorithm, multi-band remote sensing image was reduced to effective 3–4 new bands; 2) Through different threshold settings, the objects in the remote sensing image are divided into different classes; 3) The improved Otsu algorithm reduces the computational complexity, taking the threshold value of 2 as an example, the time efficiency is improved by 42.15%.