Pseudo Features-Guided Self-Training for Domain Adaptive Semantic Segmentation of Satellite Images
نویسندگان
چکیده
Semantic segmentation is a fundamental and crucial task that of great importance to real-world satellite image-based applications. Yet widely acknowledged issue occurs when applying the semantic models unseen scenery model will perform much poorer than it was applied similar training data. This phenomenon usually termed as domain shift problem. To tackle it, this article presents self-training-based unsupervised adaptation (UDA) method. Different from previous self-training approaches which focus on rectifying improving quality pseudo labels, we instead seek exploit feature-level relation among neighboring pixels structure regularize prediction adapted model. Based assumption spatial topological maintained despite impact shift, propose novel mechanism DA by exploiting local in feature space spanned teacher model, labels are generated. Quantitative experiments four different public benchmarks demonstrate proposed method can outperform other UDA methods. Besides, analytical also intuitively verify assumption. Codes be publicly available at https://github.com/zhu-xlab/PFST .
منابع مشابه
Segmentation of satellite images using Self-Organizing Maps
Remote sensing plays a key role in many domains devoted to observation of the Earth, such as land cover/use, agriculture monitoring, military battles, oceanography...etc. There are many types of acquisition systems which have different spatial, spectral, and temporal characteristics. Some of them are passive, such as Landsat, Spot, Ikonos, Quickbird, Orbview...etc. Others are active such as SAR...
متن کاملAdaptive Satellite Images Segmentation by Level Set Multiregion Competition
In this paper, we present an adaptive variational segmentation algorithm of spectral-texture regions in satellite images using level set. Satellite images contain both textured and non-textured regions, so for each region cues of spectral and texture are integrated according to their discrimination power. Motivated by Fisher-Rao’s linear discriminant analysis, two region’s weights are defined t...
متن کاملذخیره در منابع من
با ذخیره ی این منبع در منابع من، دسترسی به آن را برای استفاده های بعدی آسان تر کنید
ژورنال
عنوان ژورنال: IEEE Transactions on Geoscience and Remote Sensing
سال: 2023
ISSN: ['0196-2892', '1558-0644']
DOI: https://doi.org/10.1109/tgrs.2023.3281503