改进贝叶斯ARTMAP的迁移学习遥感影像分类算法

Transfer Learning Using Improved Bayesian ARTMAP for Remote Sensing Image Classification

遥感分类旨在从图像光谱中提取资源环境监测可用的地理信息,然而基于模式分类的图像处理技术受光谱漂移影响而缺乏历史样本重复利用的有效策略,制约着有限目标样本下遥感分类精度的提高.针对该问题,本文构建了基于改进的贝叶斯ARTMAP神经网络的迁移学习遥感影像分类算法,通过提高谐振匹配性来抑制类别扩散,利用节点的离散增量期望最大化参数更新策略,将历史遥感样本中的地物分类先验信息迁移到目标模型当中.实验结果表明本文方法能有效利用历史遥感数据弥补缺少目标训练数据的不足,相比于其他样本利用策略大幅提高遥感影像分类精度.

Remote sensing classification aims at extracting available geographic information from image spectrum for resources and environment monitoring,but due to the spectral drift effect,the lack of effective strategies on historical sample reuse for image processing technology based on pattern classification restricts remote sensing classification accuracy with limited target samples. To solve this problem,this paper proposes a transfer learning algorithm for remote sensing classification using improved Bayesian ARTMAP neural network. More productive resonance matching is used to suppress the unattractive property of category proliferation,so that the incremental expectation maximization can be introduced to update parameters adaptively. The classification prior knowledge of the historical samples is transferred to the target model. The experimental results showthat this method can effectively compensate for the lack of target training data by reusing the historical samples and significantly improve the accuracy of remote sensing image classification compared with other sample utilization strategy.