Land-use classification based on high-resolution remote sensing imagery and deep learning models.

Area of Science:

• Remote Sensing
• Artificial Intelligence
• Geospatial Analysis

Background:

• Deep learning models are crucial for land-use mapping using high-resolution imagery.
• Several new deep learning network modeling methods have emerged, but their comparative performance is unclear.

Purpose of the Study:

• To systematically compare the performance of four established deep learning models (FCN-8s, SegNet, U-Net, and Swin-UNet) for land-use mapping.
• To evaluate model generalization abilities using intersection of union and F1 scores.

Main Methods:

• Application of FCN-8s, SegNet, U-Net, and Swin-UNet models to an open benchmark high-resolution remote sensing dataset.
• Quantitative assessment of overall accuracy, intersection of union, and F1 score for each model.

Main Results:

• Swin-UNet achieved the highest overall accuracy (96.01%), followed by U-Net (91.90%), SegNet (89.86%), and FCN-8s (80.73%).
• Swin-UNet demonstrated superior robustness and generalization ability compared to the other models based on intersection of union and F1 score metrics.

Conclusions:

• Swin-UNet is the most effective deep learning model for high-resolution land-use mapping among those tested.
• The study provides a critical reference for model selection in land-use mapping, urban functional area recognition, and natural resource management.