MGML: Multigranularity Multilevel Feature Ensemble Network for Remote Sensing Scene Classification.

Area of Science:

• Computer Vision
• Geospatial Analysis
• Machine Learning

Background:

• Remote sensing (RS) scene classification faces challenges due to large intraclass variance and confusing geographic information.
• Existing methods struggle to effectively address these inherent issues in RS imagery.

Purpose of the Study:

• To propose an efficient network, the multigranularity multilevel feature ensemble network (MGML-FENet), for tackling RS scene classification.
• To enhance feature extraction and reduce interference from complex scene information.

Main Methods:

• Introduced a multigranularity multilevel feature fusion branch (MGML-FFB) using channel-separate feature generators (CS-FG) for diverse feature extraction.
• Developed a multigranularity multilevel feature ensemble module (MGML-FEM) with full-channel feature generators (FC-FG) to mitigate confusing information.
• Employed ensemble learning to combine diverse predictions for robust classification.

Main Results:

• The proposed MGML-FENet demonstrates superior performance compared to state-of-the-art methods on multiple benchmark RS datasets (AID, NWPU-RESISC45, UC-Merced, VGoogle).
• The network effectively utilizes structural information and fine-grained features.
• Visualization analysis confirmed the interpretability of the MGML-FENet model.

Conclusions:

• MGML-FENet offers a significant advancement in remote sensing scene classification by effectively addressing key challenges.
• The proposed architecture provides more convincing predictions through feature fusion and ensemble learning.
• The method shows promise for accurate and interpretable analysis of remote sensing imagery.