基于自适应共振模型的遥感影像分类方法研究

Adaptive Resonance Theory for Classification of Remotely Sensed Image

人工神经网络 (ANN)是人视觉和脑的基本功能的抽象、简化和模拟。在对遥感影像的综合解译应用中 ,与传统的统计方法和符号逻辑方法相比较 ,ANN更接近人对影像的视觉解译分析过程。自适应共振理论 (ART)是一种自组织产生认知编码的神经网络理论 ,其自组织、反馈式增量学习机能 ,能兼顾适应性和稳定性 ,克服了一般神经网络学习速度慢、网络结构难以确定、局部最小陷阱等缺陷。以FUZZY ART和ARTMAP为基础 ,提出基于ART遥感影像非监督和监督分类的一般模型 ,并以实际土地覆盖分类和城市结构信息提取为应用实例 ,通过与传统统计方法和一般ANN分类器相比较 ,ART具有正确率更高、学习速度快、自适应性等优点 ,是复杂数据分类和信息提取的有效工具。

Artificial Neural Networks (ANN) have been studied for simplified simulation to the activation of human brain and vision. Applied in understanding and interpretation of remotely sensed image, ANN can be performed more similarly with the human vision interpretability of the image in comparison with the conventional techniques such as statistical classifiers and rule-based symbolic inference. Adaptive resonance theory (ART) is developed on the basis of self-association cognitive coding theory. The major mechanism of ART is its feedback incremental learning with self-organizing structure, by which the stability and adaptability can be possessed simultaneously and the shortcomings in conventional multilayer feedforward ANN can be overcome, especially in learning phase, determination of network structure and local convergency. In this study, we firstly overview the unsupervised and supervised models of ART (including FUZZY-ART and ARTMAP), then propose a general ART-based information extraction and classification framework for remotely sensed image. With the experimental applications of land-cover classification and urban context information extraction by the presented models, the results are synthetically analyzed in comparison with conventional classifiers. The conclusion can be reached that ART model can be an alternative powerful tool for information extraction and classification of remotely sensed image especially for the features with high dimension.