Sub-pixel Unpaved Roads Detection in Landsat Images

Unpaved roads are one of the most relevant causes of erosion process in rural regions. Locating unpaved roads in low-resolution satellite images is not a trivial task. Although they have the spectral response of bare soil, their spatial width is often far below the pixel size. Therefore, the pixels crossed by unpaved roads have a different spectral response over the scene depending on the surrounding pixels. In this work, a new method to detect sub-pixel unpaved roads is proposed. The method tries to emulate the human ability to detect such roads in spite of all appointed aspects, by searching for long, narrow and smooth line segments whose spectral response are closer to the bare soil than the surrounding pixels. The method performance is evaluated on segments of Landsat scenes (bands 3, 4 and 5) of the Alcinópolis County, located in the State of South Mato Grosso in Brazil.