Connecting the dots: Triadic clustering of crowdsourced data to map dirt roads

Road segmentation is a critical application of satellite and aerial remote sensing. Traditional attempts to apply machine learning and computer vision have yielded good results but rely on specific characteristics, such as the contrast of paved roads to their surroundings or contextual clues. However, these methods still lack the sensitivity of human perception when identifying rural, non-paved, or less defined roads. We propose combining crowdsourced human labeling with triadic linear clustering to accurately map rural roads across the sparsely populated Mongolian steppe. From 600,000 road annotations made by 8000 volunteer participants we selected a random dataset to apply our proposed approach. We report the performance of the method used compared to the current state-of-the-art in automated and semi-automated satellite image road detection.