Hydrological Streamline Detection with CyberGIS-Jupyter Using Deep Learning

Author(s): 𝑍𝑒𝑀𝑒𝑖 𝑋𝑒, π‘π‘Žπ‘‘π‘‘π‘Žπ‘π‘œπ‘› π½π‘Žπ‘Ÿπ‘œπ‘’π‘›π‘β„Žπ‘Žπ‘–, π΄π‘Ÿπ‘π‘Žπ‘› π‘€π‘Žπ‘› π‘†π‘Žπ‘–π‘›π‘—π‘’, 𝐿𝑖 πΆβ„Žπ‘’π‘›,π‘β„Žπ‘–π‘¦π‘’ 𝐿𝑖, πΏπ‘Žπ‘Ÿπ‘Ÿπ‘¦ π‘†π‘‘π‘Žπ‘›π‘–π‘ π‘™π‘Žπ‘€π‘ π‘˜π‘–, πΈπ‘‘β„Žπ‘Žπ‘› π‘†β„Žπ‘Žπ‘£π‘’π‘Ÿπ‘ , 𝐡𝑖𝑛 𝑆𝑒, π‘β„Žπ‘’ π½π‘–π‘Žπ‘›π‘”, π‘†β„Žπ‘Žπ‘œπ‘€π‘’π‘› π‘Šπ‘Žπ‘›π‘”

Surface water is an irreplaceable strategic resource for human survival and social development. The accurate delineation of hydrological streamlines is critically important in various scientific disciplines, such as the assessment of present and future water resources, climate models, agriculture suitability, river dynamics, wetland inventory, watershed analysis, surface water survey and management, flood mapping, and environment monitoring. Traditional hydrologic models generate streamlines solely based on topologic information, which inevitably contain errors. For example, dried out drainage lines would always be falsely recognized as streamlines. The traditional method also ignores the information from the complex 3D environment of streamlines and surface reflectance information, which would potentially be very helpful to accurately delineate streamlines. In recent years, the availability of high accuracy LiDAR data provides us a promising method to capture both 3D information of the environment and also surface reflectance information of land cover. LiDAR sensors use NIR light in the form of a pulsed laser to measure ranges (variable distances) to the ground and also reflectance information with multiple returns. These light pulses generate precise, three-dimensional information about the shape of the surface characteristics. In this research, multiple LiDAR feature maps are generated, and we develop a U-net model for doing the streamline detection and we also test several traditional machine learning methods as our baseline for comparison. Our accuracy evaluation shows that our U-net model is able to outperform the best baseline method by 8.08% on average in F1-score and provide better smoothness and connectivity over the classified streamline channels.

Open with CyberGISX

(for viewing purpose only)

Open with CyberGISX


Name and email are required. Your email will not be published.

Please provide a username.
Please provide a valid email
Please input your message.