Weakly supervised deep learning method for building change detection
Weak labels have received great attention in deep learning-based remote sensing applications. In this talk, I will go through my previous work in the application of weak labels to deep learning-based building change detection. This talk will cover two research work with image-level and crowd-sourced labels. The first one is about building change detection using crowd-sourced labels. This study proposed a full-level fused cross-task transfer learning method and a noise-robust building extraction approach. A total of 30 scenes of ZY-3 images covering 27 major cities and corresponding crowdsourced building labels from Gaode Map were used to validate the performance of the proposed method. The second research is about built-up area change detection using image-level labels (i.e., each image is assigned to a category). This research proposed a multi-scale class activation map (CAM) extracted from image-level images and an adaptive online noise correction strategy to improve the quality of pseudo labels extracted from CAM. A total of 61 ZY-3 images covering 48 Chinese cities were used as experimental data.
About the speaker: Dr.Yinxia Cao received the Ph.D. degree from Wuhan University, China, in 2023. Now, she is a postdoc research fellow with the Department of Land Surveying and Geo-Informatics, Hong Kong Polytechnic University. Her research focuses on high- resolution remote sensing image interpretation with deep learning techniques. She has published five papers in esteemed SCI journals including Remote Sensing of Environment and ISPRS Journal of Photogrammetry and Remote Sensing. She has received the First Prize in Academic Scholarship from Wuhan University (2022), Wang Zhizhuo Innovative Talent Award (2022), and National Scholarship for Doctoral Students (2020).
Add Event to Calendar
