EXPLORING SAR-CENTERED MULTI-SOURCE DATA FUSION FOR FLOOD MAPPING WITH TRANSFORMER AND CONVOLUTION BASED ENSEMBLE MODELS

被引:0
作者
Zou, Shuchang [1 ]
Yang, Qian [1 ]
机构
[1] PIESAT Informat Technol Co Ltd, Beijing, Peoples R China
来源
IGARSS 2024-2024 IEEE INTERNATIONAL GEOSCIENCE AND REMOTE SENSING SYMPOSIUM, IGARSS 2024 | 2024年
关键词
Flood mapping; SAR-centered multi-source data fusion; Model ensemble;
D O I
10.1109/IGARSS53475.2024.10642674
中图分类号
学科分类号
摘要
The occurrence of severe hydrometeorological disasters, especially flood, has been escalating at an alarming rate. It is of great interest in exploiting the all-time and all-weather monitoring capability of SAR data to achieve rapidly and accurately flood mapping. However, the interferential speckle noises, the hill shadow and the lacking texture water body appearance in SAR data inevitably lead to false segmentation. To address these issues, we propose an effective deep learning framework to acquire flood map with SAR-centered multi-source data fusion. Specifically, we integrate both representative transformer based (Segformer) and convolution based (DCNv4+UperNet) semantic segmentation networks, thus combines both global self-attention and local inductive bias to improve the performance. Moreover, multi-band input data preprocessing, weighted model ensemble, and additional training and testing strategies are utilized to increase the segmentation accuracy. Comprehensive experiments demonstrate the effectiveness of our proposed flood mapping method which ranks in the third place of the 2024 IEEE GRSS Data Fusion Contest Track 1 test phase (F1:79.043%).
引用
收藏
页码:776 / 779
页数:4
相关论文
共 13 条
[11]  
Xie E., 2021, ADV NEURAL INF PROCE, V34, P12077, DOI DOI 10.48550/ARXIV.2105.15203
[12]  
Xiong Yuwen, 2024, ARXIV
[13]   Siam-DWENet: Flood inundation detection for SAR imagery using a cross-task transfer siamese network [J].
Zhao, Bofei ;
Sui, Haigang ;
Liu, Junyi .
INTERNATIONAL JOURNAL OF APPLIED EARTH OBSERVATION AND GEOINFORMATION, 2023, 116