A Multi-Level Feature Fusion Network for Remote Sensing Image Segmentation

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Automated scene analysis of high-resolution remote sensing images is crucial for diverse applications.
• Existing methods struggle with significant scale variations in target objects, particularly small ones.
• Applications include environmental disaster monitoring, forestry, agriculture, urban planning, and road analysis.

Purpose of the Study:

• To propose a novel network for improved high-resolution remote sensing image segmentation.
• To address the challenge of segmenting objects with large scale differences.
• To enhance the recognition accuracy of small objects in remote sensing imagery.

Main Methods:

• Development of a multi-level feature fusion network (MFNet).
• Integration of multi-level features from the network's backbone.
• Experimental validation on the Vaihingen and Potsdam datasets.

Main Results:

• The proposed MFNet achieves good segmentation results on benchmark datasets.
• The multi-level feature fusion approach effectively handles scale variations.
• Recognition performance for small objects is improved to a certain extent.

Conclusions:

• The MFNet demonstrates efficacy in high-resolution remote sensing image segmentation.
• Multi-level feature fusion is a viable strategy for scale-invariant segmentation.
• The study contributes to advancing automated scene analysis in remote sensing.