Ghost引导UNet++的高分遥感影像变化检测

Ghost-guided UNet++for high-resolution remote sensing image change detection

目的 随着遥感观测技术的飞速发展,遥感影像的分辨率越来越高,如何从高分遥感影像中有效提取具有鉴别性的特征进行地物变化检测成为一个具有挑战性的问题。卷积神经网络广泛应用于计算机视觉领域,但面向遥感影像变化检测时仍存在图像语义或位置信息的丢失及网络参数量过大等缺陷,导致检测性能受限。为此,提出一种新型GUNet++(Ghost-UNet++)网络,用于遥感影像的精准变化检测。方法 首先,为了提取双时相遥感影像更具判别性的深度特征,设计具有多分支架构的高分辨率网络HRNet替换传统UNet++的主干网;其次,采用UNet++解码结构进行差异判别时,引入鬼影(Ghost)模块代替传统卷积模块以降低网络参量,并设计密集跳跃连接进一步加强信息传输,以减少深层位置信息的丢失;最后,设计一个集成注意力模块,将网络的多个语义层次特征进行聚合和细化,抑制语义和位置信息的丢失,进一步增强特征表征能力用于最终的精准变化检测。结果 在LEVIR-CD(LEVIR change detection data set)和Google Data Set两个公开数据集上进行实验,结果表明本文算法变化检测精度高达99.62%和99.16%,且网络参数量仅为1.93 M,与现有主流变化检测方法相比优势明显。结论 提出方法综合考虑了遥感图像中语义和位置信息对变化检测性能的影响,具有良好的特征抽取和表征能力,因此变化检测的精度和效率比现有同类方法更高。

Objective With the rapid development of remote sensing observation technology,the resolutions of remote sens⁃ing images(RSIs)are increasing.Thus,how to extract discriminative features effectively from high-resolution RSIs for ground-object change detection has become a challenging problem.The existing RSI change detection methods can be divided into two categories:methods based on conventional image processing approaches and methods based on deep learn⁃ing(DL)theory.The former extracts low-level or mid-level features from RSIs for change detection,making it easy to implement and have high detection efficiency.However,the increasing resolution of RSIs result in the images having rich ground objects and complex background clutter;thus,the low-or mid-level features can hardly meet the demand of precise change detection. In recent years, DL has been introduced into the field of high-resolution RSI change detection because ofits powerful feature extraction capability. Various methods based on convolutional neural networks (CNNs) have been pro⁃posed for RSI change detection. Compared with conventional image processing methods, CNNs can extract high-levelsemantic information for high-resolution RSIs, which is beneficial to precise detection. Although CNNs have greatly raisedthe accuracy of change detection, they always involve numerous parameters and have high computational complexity. Toraise the efficiency of change detection, many scholars have proposed to perform parameter pruning on pretrained models ordesign simple network structures. However, these strategies lead to the loss of some crucial image information, includingsemantics and location information, thus reducing the detection accuracy. Therefore, this study proposes a novel GhostUNet++ (GUNet++) network for precise RSI change detection to address the problems. Method First, a high-resolutionnetwork called HRNet, which has a multibranch architecture, is designed to replace the traditional UNet++ backbone andthus extract additional discriminative deep features from bitemporal RSIs. In contrast to series structures, HRNet owns aspecial parallel architecture, which can extract additional discriminative features through multiscale feature fusion. In addi⁃tion, we choose a lightweight structure (i. e. , HRNet-W16) on the basis of a thorough analysis of various existing HRNetstructures to ensure that the whole network possesses low complexity. Second, when applying the UNet++ decoding struc⁃ture for difference discrimination, the Ghost module is introduced to replace the conventional convolutional module andthus reduce the network parameters;meanwhile, a dense skip connection is designed to enhance the information transmis⁃sion further and reduce the loss of location information. The core idea of the Ghost module is to adopt simple linear opera⁃tions instead of the traditional convolutional operations to generate Ghost maps for original features, which may save sub⁃stantial computational cost. Third, an ensemble attention module is constructed to aggregate and refine the multilevelsemantic features of the network, thereby suppressing the loss of semantic and location information and further enhancingthe feature representation ability for final accurate change detection. Features generated at various levels usually containdifferent meanings: shallow ones always contain detailed spatial information, while deep ones reflect rich semantic con⁃tent. On this basis, we propose an adaptive channel selection mechanism to integrate these different features effectively.Finally, we propose to combine two different loss functions, i. e. , the sigmoid loss function and the dice loss function, forthe whole model training to enhance the detection performance further. Compared with the methods that merely use one lossfunction, this scheme can improve detection performance. Result A series of experiments is conducted on two publiclyavailable datasets, including LEVIR-CD and Google DataSet, to validate the effectiveness of the proposed method. Theexperiments consist of ablation analysis and comparison experiments. Specifically, three kinds of ablation analyses are per⁃formed. The first one verifies the effects of the modified HRNet-based discriminative feature extraction, the second onemainly evaluates the effectiveness of the ensemble attention modules and the Ghost modules for the whole network, and thethird one aims to find the optimal values of a key parameter in the Ghost models. In the comparison experiments, somestate-of-the-art algorithms are selected for comparison to verify the superiority of the proposed network. Extensive experi⁃mental results demonstrate that for the two famous datasets, the proposed method achieves high change detection accuracyrates of 99. 62% and 99. 16%. In addition, the parameter of the network is only 1. 93 M. Compared with some mainstreamchange detection approaches, the proposed method is remarkably superior. Conclusion The proposed method comprehen⁃sively considers the effect of semantic and location information in RSIs on the performance of change detection. In addi⁃tion, the method possesses good feature extraction and representation capabilities. Therefore, the accuracy and efficiencyfor change detection are higher than those of existing comparable algorithms. In the future, we plan to optimize the pro⁃posed architecture by increasing the number and diversity of the training samples to enhance the robustness of models, byusing more advanced software and hardware environments for experiments to reduce the training time, and by applying ourtrained model to other tasks.