Landsat-8影像的LDA模型变化检测

Change Detection and Analysis of Landsat-8 Image Based on LDA Model

变化检测一直是遥感研究领域的热点,随着遥感技术的不断发展,新型数据源不断涌现,使传统遥感变化检测方法面临新的挑战。本文以Landsat-8影像为主要数据源,使用影像分割算法,设计2期遥感影像的文档-单词映射,将影像中所有的像元作为视觉单词,利用LDA模型将影像文档从单词空间转换到主题空间进行表达。在此基础上,结合实地调查对变化区域进行检测和验证,形成一套面向对象的LDA模型变化检测方法。研究表明:基于图斑的分析可有效消除以像元尺度进行变化检测产生的椒盐现象;利用LDA模型构建的变化检测方法能较好地实现影像文档特征的统一表达,有效去除2期影像相同地物因光谱差异导致的变化误检验;与差值法和波谱角等常规遥感变化检测方法相比,该方法能有效地减少错漏判,提高遥感影像变化检测的正确率,为中高分辨率遥感影像的变化检测提供新思路。

Change detection with remote sensing images plays an important role in land cover mapping. With the development of science and technology, a series of new remote sensing data sources have become available, and have been significantly improved, which also brings a great challenge to the traditional remote sensing change detection methods. Unlike the other traditional methods for change detection, the present work uses Latent Dirichlet Allocation model （LDA） in learning middle-level semantic topics instead of low-level features from re- mote sensing images. In this paper, we use the pixels of two remote sensing images as the basic unit, while the image segments are used as the documents in the object-based image analysis methods. Firstly, we try to extract some features from these remote sensing images, such as the spectral and textural features. Then, we work on or- ganizing the local features from these two images to obtain visual words and construct the bag of words model （BOWM） representation. Based on this, the LDA model is utilized to reveal the underlying topics, which are used to detect the change of the study area. Every document of remote sensing images has a specific topic distri- bution, which is related to the reference data of the study area. In this process, the pseudo changes and actual changes of these two remote sensing images can be distinguished by the topic distributions of the documents. Compared with traditional pixel-level change detection methods, the method of LDA-based model is less influ- enced by the spectral variance of two images, which avoids the ＂salt and pepper＂ effect by using object-based analysis method. The effectiveness of LDA-based model change detection approach was verified in experiments with the accuracy to be 85.35%, and it is also compared with techniques using Spectral Angle Mapper and Image differencing. The result shows that our studies provide a good approach to improve the accuracy and reduce the mistake rate of change detection between two images. Our work indicates that LDA model-based approach is superior to the traditional methods and the proposed method is applicable to the analysis of change area detection using Landsat-8 images.