Unsupervised Change Detection Driven by Floating References: A Pattern Analysis Approach

Abstract

The Earth's environment is continually changing due to both human and natural factors. Timely identification of the location and kind of change is of paramount importance in several areas of application. Because of that, Remote Sensing change detection is a topic of great interest. The development of precise change detection methods is a constant challenge. This study introduces a novel unsupervised change detection method based on data clustering and optimization. The proposal is less dependent on radiometric normalization than classical approaches. We carried experiments with remote sensing images and simulated data sets to compare the proposed method with other unsupervised well-known techniques. At its best, the proposal improves by 50 % the accuracy concerning the second best technique. Such improvement is most noticeable with uncalibrated data. Experiments with simulated data reveal that the proposal is better than all other compared methods at any practical significance level. The results show the potential of the proposed method.

Unsupervised Multitemporal Multiclass Change Detection

“…To achieve this, a threshold η is applied to γ(w s ) values. The empirical rule presented by Negri and Frery [36] is an alternative to classic algorithms [37,38] effective when the data follow a positive and heavy-tailed distribution, as usually observed in deviation measures such as those expressed by V. This rule is defined by:…”

Wavelet Spatio-Temporal Change Detection on Multitemporal SAR Images

“…Change detection analysis is classified either as supervised (training data is used to set up the method) or unsupervised (fully data-driven techniques). We focus here on unsupervised approaches [50]- [55]. Most of these methods have been developed for pairs of images.…”