Road Extraction - Global

This deep learning model is used to extract roads from high resolution (1 meter) aerial/satellite imagery. Road layers are useful in preparing base maps and analysis workflows for urban planning and development, change detection, infrastructure planning, and a variety of other applications.

Digitizing roads from imagery is a time-consuming task and is commonly done by digitizing features manually. Deep learning models are highly capable of learning these complex semantics and can produce superior results. Use this deep learning model to automate this process and reduce the time and effort required for acquiring road layers.

Using the model

This model can be used with Detect Objects Using Deep Learning tool in ArcGIS Pro version 3.3 or later. If you are using ArcGIS Pro versions between 2.9 to 3.2, use this alternate model. Follow the guide to use the model. Before using this model, ensure that the supported deep learning libraries are installed. For more details, check Deep Learning Libraries Installer for ArcGIS.

Fine-tuning the model
This model cannot be fine-tuned using ArcGIS tools.

Input
8-bit, 3-band high resolution (1 meter) aerial/satellite imagery.

Output
Feature class representing road network. If you also wish you get a raster containing the probability of roads at each pixel, use this variant of the model.

Applicable geographies
The model is expected to work well globally.

Model architecture
The implementation is based on the Segment Anything Model for Road Extraction by Congrui Hetang et al. 

Accuracy metrics
The model has an F1 score of 77.23 on city-scale dataset. It has a precision of 0.904 and recall of 0.683. 

Sample results
Here are a few results from the model.