Deep Feature Aggregation Framework Driven by Graph Convolutional Network for Scene Classification in Remote Sensing.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• High spatial resolution (HSR) scene classification is vital for land planning and utilization.
• Convolutional Neural Networks (CNNs) excel at feature learning but struggle with context relationships in HSR images.
• Graph-based deep learning offers powerful data relevance representation for intrinsic attribute learning.

Purpose of the Study:

• To develop a deep feature aggregation framework driven by graph convolutional network (DFAGCN) for improved HSR scene classification.
• To effectively capture patch-to-patch correlations in convolutional feature maps.
• To integrate multilayer features for enhanced classification accuracy.

Main Methods:

• Utilized an off-the-shelf CNN pretrained on ImageNet for multilayer feature extraction.
• Employed a graph convolutional network (GCN) to model patch-to-patch correlations within feature maps.
• Integrated multilayer convolutional and fully connected features using a weighted concatenation method.

Main Results:

• The proposed DFAGCN framework demonstrated competitive performance on UCM, AID, RSSCN7, and NWPU-RESISC45 datasets.
• DFAGCN achieved superior overall accuracy (OAs) compared to state-of-the-art scene classification methods.
• The GCN component effectively revealed and utilized contextual relationships in HSR images.

Conclusions:

• The DFAGCN framework offers a robust approach for HSR scene classification by effectively integrating deep features.
• Graph convolutional networks significantly enhance the ability to capture contextual information in remote sensing imagery.
• The proposed method provides a valuable advancement for practical applications requiring accurate HSR scene classification.