与传统统计方法的分类器相比较,人工神经网络(ANN)方法应用于遥感影像分类,不需预先假设样本空间的参数化统计分布,具有复杂的映射能力.大多数ANN分类器采用误差反向传播(BP)学习算法的多层感知器模型(BPNN),其主要缺陷是学习速度缓慢、容易陷入局部极小而导致难以收敛等.基于径向基函数(RBF)映射理论的神经网络模型融合了参数化统计分布模型和非参数化线性感知器映射模型的优点,在实现快速学习的同时,保持了高度复杂的映射能力.该文主要探讨RBF映射理论在遥感影像分类中的具体算法和实现过程,并初步提出了融合地学知识的RBF影像分类模型;最后以实际的遥感土地覆盖分类为例,通过与BP神经网络方法(BPNN)相比较,对分类过程和结果进行了综合分析,认为RBF方法在学习速度、网络结构、融合领域知识等方面具有一定的优势.

In recent years the artificial neural network has been developed and applied to remotely sensed data classification problem. Most modal of them are error back-propagation(BP), BP learning algorithm based multi-layer perceptron. Compared to the conventional statistical classifier, BPNN RS image classifier are non-parametric and may have the capacity of more robust proximity especially when distributions are strongly non-Gaussian, but its main shortcoming is its slow training speed, local minimum and even being unable to converge. The Radial Basis Functions Neural Network (RBFNN) modal, integrating the parametric statistic distribution modal and non-parametric single layer perceptron modal, trains faster and more stable than BPNN while keeping the complicated proximity. In this article, the survey and analysis of the RBFNN for the classification of remotely-sensed multi-spectral image is presented, and the RBF RS image classification modal, detailed algorithms and realization procedures is intially raised. The framework which fuses Geo-Knowledge into RBFNN by RBF functions and hierarchical clustering means with optimization evolution theory also are introduced. Finally, the case of practical application of remote sensing land cover classification in Hong Kong region is presented. After the procedure of RBFNN and BPNN approaches are synthetically analyzed, experimental results show that RBFNN approach has more advantages in train time, network structure, knowledge fusion, etc.