• 主题
高级搜索  >
历史搜索 清空历史
首页> 外文会议>Asian conference on remote sensing;ACRS >Domain Adaptation for 2D/3D Change Detection in VHR Imagery via Calibration of Convolutional Neural Network under Prior Probability Shift
【24h】

Domain Adaptation for 2D/3D Change Detection in VHR Imagery via Calibration of Convolutional Neural Network under Prior Probability Shift

机译：在先验概率平移下通过卷积神经网络校准在VHR图像中进行2D / 3D变化检测的域自适应

获取原文
页面导航
  • 摘要
  • 著录项
  • 相似文献
  • 相关主题

摘要

The success of deep learning algorithms applied on large-scale remotely sensed imagery rests on the representativeness of the samples used to train the model. When the source domain training samples do not fully represent the target domain samples where no prior information is available, the trained model may fail to adapt to the target domain, which makes it compulsory to repeat the time-consuming labeling process. Even though this issue could be a critical point to apply deep learning algorithms to change detection in remote sensing, the vast majority of change detection algorithms have not considered the discrepancy between the two domains, and in turn, failed to generalize their performance on a newly given area. To address this problem, we focus on the discrepancy where the class distributions of the two domains are not balanced, defined as the prior probability shift, associated with the supervised change detection. This research shows that the performance of a convolutional neural network (CNN) is largely impaired by the inconsistent changed ratio and proposes a novel approach to resolve the problem of prior probability shift. The proposed framework estimated the changed ratio of the target domain by patch-based change vector analysis (CVA) and calibrated the change threshold from the softmax output of the CNN to adapt well to the target domain. The framework was implemented in three sub-regions with various changed ratios of 20.0 %, 45.0 %, and 35.0 %, acquired from bi-temporal Unmanned Aerial Vehicle (UAV) imagery with 0.1-m resolution including RGB channels and digital surface model (DSM). A total of 1,000 and 2.000 training samples drawn from each of the three sub-regions (source domain), with the two different sampling scenarios (i.e., balanced sampling and imbalanced sampling), were trained by a CNN. Subsequently, the trained CNN was adapted using the patch-based CVA and tested on the two other sub-regions (target domains) of different changeon-change ratios, respectively. The proposed framework was separately applied to RGB imagery and RGB + DSM imagery. The results demonstrated an improvement in the change detection performance by evaluating the overall accuracy, recall, precision, and Fl-score, thus confirming the effectiveness of the proposed framework to address the prior probability shift problem.
机译：应用于大型遥感影像的深度学习算法的成功取决于用于训练模型的样本的代表性。当源域训练样本不能完全代表没有可用先验信息的目标域样本时，训练后的模型可能无法适应目标域，这使得必须重复耗时的标记过程。尽管此问题可能是将深度学习算法应用于遥感中的变化检测的关键点，但绝大多数变化检测算法并未考虑这两个领域之间的差异，因此未能在新的领域中概括其性能。给定区域。为了解决这个问题，我们集中在两个域的类分布不平衡的差异上，该差异定义为与有监督的更改检测相关联的先验概率偏移。这项研究表明，卷积神经网络（CNN）的性能在很大程度上受可变比率变化的影响，并提出了一种解决先验概率转移问题的新方法。拟议的框架通过基于补丁的更改矢量分析（CVA）估算了目标域的变化比例，并根据CNN的softmax输出校准了变化阈值，以很好地适应目标域。该框架是在三个子区域中实施的，其变化率分别为20.0％，45.0％和35.0％，这些区域是从具有0.1微米分辨率的双时空无人机（UAV）图像获取的，包括RGB通道和数字表面模型（DSM） ）。有线电视新闻网（CNN）分别从三个子区域（源域）中抽取了1,000和2.000个训练样本，并使用了两种不同的采样方案（即平衡采样和不平衡采样）。随后，使用基于补丁的CVA对经过训练的CNN进行调整，并分别在其他两个具有不同更改/未更改比率的子区域（目标域）上进行测试。所提出的框架分别应用于RGB图像和RGB + DSM图像。结果表明，通过评估整体准确性，查全率，准确性和F1得分，可以改善变更检测性能，从而证实了所提出框架解决先前概率转移问题的有效性。

著录项

相似文献

  • 外文文献
  • 中文文献
  • 专利
获取原文

客服邮箱：kefu@zhangqiaokeyan.com

京公网安备：11010802029741号 ICP备案号：京ICP备15016152号-6 六维联合信息科技 (北京) 有限公司©版权所有

  • 客服微信

  • 服务号