融合面向对象与卷积神经网络的GF-2古城墙提取技术分析

Analysis of GF-2 Ancient City Wall Extraction Technology Based on Object-oriented Combined with Convolutional Neural Network

针对目前由于地形与气候等限制条件导致对一些古建筑遗址很难进行动态监测和保护的问题,提出基于高分影像,采用面向对象结合卷积神经网络分类算法实现对大型土筑古城武威满城城墙的精细化提取,并与最大熵耦合离散粒子群算法(MEDPSO)及最大似然法(MLC)对比,验证该方法的适用性和精度。提取结果表明:面向对象结合卷积神经网络分类方法表现出很强的抗干扰和泛化能力,能够实现对城墙边界的有效提取。本文方法的Kappa系数(0.95)高于MEDPSO(0.92)和MLC(0.86);其总体精度(97.46%)高于MEDPSO(95.68%)和MLC(92.67%),从而验证了提出方法对古建筑城墙提取的有效性,为古城墙信息提取提供技术参考和借鉴价值。

In view of the current problem that it is difficult to dynamically monitor and protect some ancient architectural sites due to the limitations of terrain and climate, an object-oriented classification algorithm combined with convolutional neural network is proposed based on high-resolution images to realize the fine extraction of the walls of the large earthen ancient city of Wuweiman. Compared with the maximum entropy coupled discrete particle swarm algorithm(MEDPSO) and maximum likelihood classification(MLC), the applicability and accuracy of the method are verified. The results show that the object-oriented classification method combined with convolutional neural network shows strong anti-interference and generalization ability and can effectively extract the boundary of the city wall. Its Kappa(0.95) is higher than those of MEDPSO(0.92) and MLC(0.86). And its overall accuracy(97.46%) is higher than those of MEDPSO(95.68%) and MLC(92.67%). The effectiveness of the object-oriented classification method combined with convolutional neural network for the extraction of ancient building walls is verified, which provides technical reference and reference value for the extraction of ancient city wall information.