Enhancing Land Cover Classification Accuracy Using Modified U-Net Architecture With SK-ResNeXt as the Encoder
Researchers have developed an innovative method for land cover classification by integrating SK-ResNeXt with U-Net, demonstrating notable improvements using multispectral imagery.
The study investigates enhancements to land cover classification (LCC) using U-Net and SK-ResNeXt, achieving accuracy gains of over 5% compared to baseline models. The researchers focused on land cover types, which are increasingly important for urban planning, ecological conservation, and resource management. Through this approach, they hope to address the limitations of traditional methods.
This integrated model improves feature extraction and processing across different scales—capabilities especially valuable when dealing with complex datasets like the Five-Billion-Pixels dataset, released in 2023. This dataset comprises over 5 billion labeled pixels spanning 150 high-resolution RGB-NIR images across more than 60 administrative districts across China.
The modified U-Net architecture utilizes SK-ResNeXt as its encoder, which empowers the model to capture diverse spatial features and adapt to variations across the dataset. This enhancement effectively balances both accuracy and computational efficiency. The outcomes reveal how introducing diverse spectral bands leads to significantly improved discrimination of complex land cover types.
Throughout testing, models employing this integrated architecture demonstrated superior performance compared to venerable models like DeepLabV3 and PSPNet, particularly on nuanced classes such as lakes and vegetation.
The findings not only validate the methodological framework employed but also illuminate prospects for future research, where even finer adjustments to the U-Net architecture could yield even richer insights and applications for land cover classification.