BP神经网络已被证明能有效实现遥感图像的亚像元定位,但其训练时间较长,容易陷入局部最优且依赖于大量的训练样本。而在实际应用中,训练样本即先验信息较难获取。然而,建筑物及道路等地物具有规则的线性空间分布。针对这些线性特征地物,研究了一种训练样本的几何合成方法,消除对先验信息的依赖,并提出利用最小二乘支持向量机实现亚像元定位。实验表明,这种结合合成训练样本与最小二乘支持向量机的亚像元定位方法是合理可行的,且与BP神经网络实现方法相比,训练过程明显加快,定位精度更高。 BP neural network has been proved to be effective in remote sensing image sub-pixel localization, but its training time is long, easy to fall into the local optimum and depends on a large number of training samples. However, in practice, training samples, that is, prior information, are difficult to obtain. However, objects such as buildings and roads have a regular linear spatial distribution. For these linear features, a geometric synthesis method of training samples is studied, which eliminates the dependence on prior information and proposes the use of least square support vector machine to achieve sub-pixel localization. Experiments show that this method of combining sub-pixel positioning with the combination of training samples and LS-SVM is reasonable and feasible. Compared with the BP neural network, the training process is obviously accelerated and the positioning accuracy is higher.