1.中国地质大学（武汉）自动化学院，武汉 430074;2.复杂系统先进控制与智能自动化湖北省重点实验室，武汉 430074;3.地球探测智能化技术教育部工程研究中心，武汉 430074
1.School of Automation, China University of Geosciences, Wuhan 430074, P. R. China;2.Hubei Key Laboratory of Advanced Control and Intelligent Automation for Complex Systems, Wuhan 430074, P. R. China;3.Engineering Research Center of Intelligent Geodetection Technology Ministry of Education, Wuhan 430074, P. R. China
the Natural Science Foundation of Hubei Province, China, under Grant 2020CFA031
遥感图像场景分类对土地资源管理具有重要意义, 然而高分辨率遥感图像中地物分布复杂, 图像中存在着与当前场景无关的冗余信息, 会对场景的精确分类造成影响. 针对该问题, 提出一种基于脉冲卷积神经网络(Spike Convolutional Neural Network, SCNN) 稀疏表征的场景分类方法. 从稀疏表征出发, 利用脉冲神经元的稀疏脉冲输出特性, 设计脉冲卷积神经网络, 去除遥感图像中与场景无关的冗余信息, 实现对图像的稀疏表征; 提出了基于脉冲输出交叉熵损失函数的反向传播算法, 在该算法的基础上利用梯度下降训练脉冲卷积神经网络, 优化网络参数, 实现遥感图像场景分类. 通过实验验证方法的有效性, 所提方法应用于 Google 和 UCM 两个遥感图像数据集, 并与传统的卷积神经网络 (Convolutional Neural Network, CNN) 进行了对比. 实验结果表明, 所提方法可以对遥感图像进行稀疏表征, 实现场景分类; 而且相对于卷积神经网络, 所提方法在遥感图像场景分类任务上更有优势.
Remote sensing imagery scene classification is of great significance to land resource management. However, the distribution of ground objects in remote sensing images of High Spatial Resolution (HSR) is complex, and there is redundant information irrelevant to the current scene in the images, which will affect the accurate classification of the scene. To solve this problem, a scene classification method based on spike convolutional neural network (SCNN) is proposed. From the perspective of sparse representation, the SCNN is designed based on the sparse spike output characteristics of spike neurons to remove the redundant information irrelevant to the scene in remote sensing images and realize sparse representation of images. A backpropagation algorithm based on the spike output cross entropy loss function is proposed. Based on this algorithm, the SCNN is trained by gradient descent, and the network parameters are optimized to realize scene classification of remote sensing images. The validity of the proposed method was verified by experiments, where the proposed method was applied to two remote sensing imagery datasets, namely, Google and UCM, and compared with the traditional convolutional neural network (CNN). Experimental results demonstrated that the proposed method was able to perform sparse representation of remote sensing images and realize scene classification; and compared with CNN, the proposed method showed better performance in remote sensing imagery scene classification task.