Comparing UNet configurations for anthropogenic geomorphic feature extraction from land surface parameters

Area of Science:

• Geospatial analysis
• Deep learning
• Computer vision

Background:

• Deep learning, particularly UNet, excels in semantic segmentation for geospatial and medical imaging.
• Anthropogenic geomorphic feature extraction faces challenges like small datasets and class imbalance.
• Modifications to UNet are explored to enhance performance in these specific tasks.

Purpose of the Study:

• Investigate architectural UNet modifications for improved geomorphic feature extraction.
• Evaluate performance across various training sample sizes.
• Determine the efficacy of advanced modules versus base UNet.

Main Methods:

• Implemented modifications to UNet: activation functions (leaky ReLU, swish), residual connections, squeeze and excitation modules, attention gates, dilated convolutions, and MobileNetV2 backbone.
• Utilized unique geomorphic datasets from high-resolution lidar data.
• Tested models on mapping agricultural terraces, mine benches, and valley fill faces with varying training data (50-full set).

Main Results:

• Advanced UNet modules enhanced segmentation performance, especially with limited training data and complex landscapes.
• Performance gains diminished with larger training sets (above 500 chips).
• Base UNet architecture proved adequate for many scenarios, particularly with sufficient data.

Conclusions:

• UNet serves as a flexible framework adaptable for geomorphic feature extraction.
• Optimized UNet models offer valuable insights for improving deep learning in geospatial applications.
• Further research can build on these findings for more accurate and efficient feature mapping.