A Novel Adaptive Hybrid Fusion Network for Multiresolution Remote Sensing Images Classification
Autor: | Wenping Ma, Jun Zhang, Licheng Jiao, Biao Hou, Jiliang Zhao, Hao Zhu, Jianchao Shen |
---|---|
Rok vydání: | 2022 |
Předmět: |
Backbone network
Contextual image classification Channel (digital image) Computer science Feature (computer vision) Multispectral image ComputingMethodologies_IMAGEPROCESSINGANDCOMPUTERVISION General Earth and Planetary Sciences Pyramid (image processing) Electrical and Electronic Engineering Sensor fusion Panchromatic film Remote sensing |
Zdroj: | IEEE Transactions on Geoscience and Remote Sensing. 60:1-17 |
ISSN: | 1558-0644 0196-2892 |
Popis: | With the rapid development of earth observation technology, panchromatic (PAN) and multispectral (MS) images have also become easier to obtain. The multiresolution classification of PAN and MS images as a basic MS image analysis task has become a research hotspot. The main challenge in this field is how to process data and extract features to improve classification accuracy effectively. In this article, we design a novel adaptive hybrid fusion network (AHF-Net) for multiresolution remote sensing image classification. It includes two parts: data fusion and feature fusion. In the data fusion part, we propose an adaptive weighted intensity-hue-saturation (AWIHS) strategy, which can reduce the difference between MS and PAN images by adaptively adding each other's unique information from the perspective of information sharing. In the feature fusion part, starting from the second-order correlation of features, we propose a correlation-based attention feature fusion (CAFF) module. It can improve the discrimination of fusion features by adaptively determining the fusion coefficient according to the importance of the input feature channel. Based on AWIHS and CAFF, inspired by the idea of feature pyramid, we combine the multilevel feature fusion and the dual-branch residual network as the backbone network of AHF-Net. By combining AWIHS and CAFF modules with the backbone network, our AHF-Net can effectively improve the classification accuracy of multiresolution remote sensing images. The effectiveness of the proposed algorithm has been verified on multiple data sets. Our code and model are available at https://github.com/1826133674/AHF-Net. |
Databáze: | OpenAIRE |
Externí odkaz: |