SKUNet: Leveraging U-Net and Selective Feature Extraction for Land Cover Classification Using Remote Sensing Imagery

• Official implementation and model code are available
• The SKUNet paper is available here
• SKUNet has been accepted in Scientific Reports

SKUNet is a U-Net-based semantic segmentation model for land-cover classification using multispectral remote sensing imagery. It enhances the conventional U-Net architecture with a Selective Kernel ResNeXt50 encoder, combining the multi-path representation capacity of ResNeXt with adaptive receptive fields that dynamically capture features at different spatial scales. This design improves multiscale feature extraction while preserving the spatial reconstruction capabilities of the U-Net decoder. The model supports multiband inputs and is evaluated on the Five-Billion-Pixels dataset, which contains 24 land-cover classes on RGB, RG-NIR, and RGB-NIR configurations.

pip install -r requirements.txt

SKUNet can be instantiated directly from skunet.py and integrated into any training workflow:

from skunet import SKUNet

model = SKUNet(
    in_channels=4,
    num_classes=24,
)

• Input: tensor of shape (B, in_channels, H, W)
• Output: logits of shape (B, num_classes, H, W)

Example forward pass:

import torch
from skunet import SKUNet

model = SKUNet(in_channels=4, num_classes=24)
x = torch.randn(2, 4, 256, 256)
y = model(x)
print(y.shape)  # (2, 24, 256, 256)

SKUNet is evaluated on the Five-Billion-Pixels dataset, a large-scale land-cover classification benchmark containing:

• 150 high-resolution Gaofen-2 satellite images
• More than 5 billion annotated pixels
• 24 land-cover classes
• Four spectral bands: blue, green, red, and near-infrared
• 4 m spatial resolution

The original large tiles are cropped into non-overlapping 256 x 256 patches. The model is evaluated using three band configurations:

• RGB
• RG-NIR
• RGB-NIR

• SK-ResNeXt50 encoder with adaptive multi-scale feature extraction.
• U-Net decoder with spatial skip connections for precise reconstruction.
• Support for RGB and multispectral input configurations.
• Pretrained encoder for stronger feature initialization.
• Competitive accuracy on a large-scale, 24-class land-cover benchmark.

If you find this work useful, please star ⭐️⭐️⭐️ our repository and cite our paper:

@article{ramos2025leveraging,
    title = {Leveraging U-Net and selective feature extraction for land cover classification using remote sensing imagery},
    volume = {15},
    number = {1},
    journal = {Scientific Reports},
    author = {Ramos, Leo Thomas and Sappa, Angel D.},
    year = {2025},
    pages = {784},
    doi = {10.1038/s41598-024-84795-1}
}

Distributed under the GNU General Public License v3.0. See LICENSE for more information.