Leveraging U-Net and selective feature extraction for land cover classification using remote sensing imagery.

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Land Cover Classification (LCC) is crucial for environmental monitoring.
• Traditional U-Net architectures face challenges with multi-scale features and spatial resolution variations in Multispectral Imaging (MSI).

Purpose of the Study:

• To enhance the U-Net architecture for improved LCC using MSI.
• To integrate SK-ResNeXt as an encoder to capture multi-scale features and adapt to spatial resolution variations.

Main Methods:

• The study proposes integrating SK-ResNeXt with U-Net for LCC.
• The Five-Billion-Pixels dataset (150 RGB-NIR images, 5 billion pixels, 24 categories) was used for evaluation.
• Performance was compared against baseline U-Net and other models like DeepLabV3, SegFormer, and PSPNet.

Main Results:

• The SK-ResNeXt-enhanced U-Net achieved significant improvements in Overall Accuracy (OA) and mean Intersection over Union (mIoU) across RGB, RG-NIR, and RGB-NIR configurations.
• Improvements ranged from 5.312% to 6.928% in OA and 8.906% to 6.938% in mIoU compared to baseline U-Net.
• The enhanced model outperformed established and state-of-the-art methods, particularly in segmenting challenging classes like water bodies and industrial areas.

Conclusions:

• The integration of SK-ResNeXt effectively enhances U-Net for MSI-based LCC.
• The proposed approach demonstrates superior performance in segmenting complex land cover types.
• This method offers a promising solution for accurate and robust land cover mapping.