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

Area of Science:

• Geospatial analysis
• Deep learning
• Geomorphology

Background:

• Deep learning, particularly UNet, excels in semantic segmentation for image analysis.
• Extracting anthropogenic geomorphic features presents challenges like small sample sizes and class imbalance.
• Modifications to UNet are explored to enhance performance in these specific tasks.

Purpose of the Study:

• To investigate various architectural modifications to the UNet model for improved geomorphic feature extraction.
• To evaluate the performance of modified UNet architectures using high-resolution lidar data.
• To analyze the impact of different training sample sizes on model performance.

Main Methods:

• Implemented modifications to the base UNet architecture, including activation function changes (leaky ReLU, swish), residual connections, squeeze and excitation modules, attention gates, dilated convolutions, and MobileNetV2 backbone.
• Utilized unique geomorphic datasets from high spatial resolution lidar data for mapping agricultural terraces, mine benches, and valley fill faces.
• Assessed model performance across diverse training sample sizes (50 to full dataset).

Main Results:

• Incorporating advanced modules enhanced segmentation performance, particularly with limited training data and complex landscapes.
• Performance differences diminished with larger training set sizes (above 500 image chips).
• The base UNet architecture proved adequate for most tasks when sufficient data was available.

Conclusions:

• Modified UNet architectures offer improvements for anthropogenic geomorphic feature extraction, especially under data constraints.
• UNet serves as a flexible framework adaptable to various geospatial applications.
• Further research can build upon these findings to enhance deep learning accuracy and efficiency in geospatial analysis.