LKAFFNet: A Novel Large-Kernel Attention Feature Fusion Network for Land Cover Segmentation

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Accurate land cover segmentation is vital for urban planning, environmental monitoring, and disaster management.
• Traditional Convolutional Neural Networks (CNNs) face challenges in integrating local details with large-scale context in high-resolution imagery.

Purpose of the Study:

• To develop a novel framework, LKAFFNet, that enhances land cover segmentation by addressing the limitations of traditional CNNs.
• To improve the balance between fine-grained local features and broad contextual information in remote sensing image analysis.

Main Methods:

• Introduced LKAFFNet, a framework combining large-kernel convolutions, attention mechanisms, and multi-scale feature fusion.
• Developed three key modules: LkResNet for enhanced feature extraction with large-kernel convolutions, Large-Kernel Attention Aggregation (LKAA) for integrated spatial and channel attention, and Channel Difference Features Shift Fusion (CDFSF) for efficient multi-scale fusion.

Main Results:

• LKAFFNet demonstrated superior performance compared to previous models on the LandCover and WHU Building datasets.
• Achieved a mean Intersection over Union (mIoU) of 0.8155 on the LandCover dataset and 0.9326 on the WHU Building dataset.
• The framework showed particular effectiveness in segmenting land cover across diverse scales.

Conclusions:

• LKAFFNet significantly advances land cover segmentation accuracy in high-resolution remote sensing imagery.
• The proposed framework offers a more effective tool for various remote sensing applications requiring precise land cover classification.