Road Segmentation

Model by CrowdAI

CrowdAI’s Road Segmentation model uses cutting-edge deep learning and image segmentation techniques to precisely locate and label roads on Maxar satellite imagery. Detections are output as a geoTIFF mask, with some post-processing techniques used to smooth the output and remove false positives. This model is globally flexible in line with Maxar’s coverage.

  • Description

    Product Description


    72% F1 Score

    99% Pixelwise Accuracy

    73% Precision

    73% Recall

    In the current version of CrowdAI Road Detector, the algorithm provides best results for paved roads. However, unpaved and dirt roads are often detected, but at a lower accuracy that may result in those road types being disconnected from a complete road network.

    F1 Score is the harmonic mean of the precision and recall, with best value of 1. It measures the balance between the two metrics. Further information here.

    Pixelwise accuracy measures the percent of pixels in a predicted area that are classified or predicted correctly. Further information here.

    A higher precision score indicates that the majority of labels predicted by the model for different classes are accurate. Further information here.

    A higher recall score indicates that the model finds and predicts correct labels for the majority of the classes it is supposed to find. Further information here.


    This model analyzes an input image to identify roads. A neural network is used to assign a class to every pixel: either “road” or “non-road”. Some post-processing techniques are used to smooth the output and remove false positives. A geoTiff color overlay is then created.

    Note that this algorithm does not consider the following features to be roads: Footpaths, Bike paths, Driveways, Parking lots, Waterways, Railroads, Paths in golf courses, Deforestation paths, Ports, Beaches

    This algorithm is compatible with Maxar imagery that meets the following criteria:

    • Compatible bands & sensors: 8-band or 4-band RGBN imagery: WorldView-2, WorldView-3, WorldView-4, GeoEye-1

    • 50cm/pixel orthorectified imagery

    • Pansharpening ON, Dynamic Range Adjustment (DRA) OFF, Atmospheric Compensation (AComp) ON


    This model was trained on over 25,000 image chips of Maxar 50cm orthorectified imagery across a wide variety of nations/states, geographies, biomes, seasons, and climates on all populated continents. The training data included a mix of urban, suburban, and rural geographies to maximize flexibility.

    • Suitable locations: All geographies except areas of abundant snow cover. The algorithm performs best in urban and suburban areas (without tall buildings obscuring roads), but also works in rural/remote areas. Top countries represented in the training data are (in no particular order): China, Bolivia, Japan, Mexico, Indonesia, and United States of America.

    • Suitable seasons: All, but presence of abundant snow cover in fall and winter seasons should be avoided.

    • Biomes: The algorithm has been trained on a wide variety of biomes, with temperate forests, tropical/subtropical forests, deserts/shrublands, and grasslands representing the largest categories.


    This model was validated against CrowdAI’s internal Maxar validation set, a globalized set of imagery drawing from the WorldView and GeoEye constellations.