Mapping streams and ditches using Aerial Laser Scanning

Streams and ditches are seldom identified on current maps due to their small dimensions and sometimes intermittent nature. Estimates point out that only 9% of all ditches are currently mapped, and the underestimation of natural streams is a global issue. Ditches have been dug in European boreal forests and some parts of North America to drain wetlands and increase forest production, consequently boosting the availability of cultivable land and a national-scale landscape modification. Target 6.6 of the Agenda 2030 highlights the importance of protecting and restoring water-related ecosystems. Wetlands are a substantial part of this, having a high carbon storage capability, the property of mitigating floods, and purifying water. All things accounted for, the withdrawal of anthropogenic environment alterations can be on the horizon, even more because ditches are also strong emitters of methane and other greenhouse gases due to their anoxic water and sediment accumulation. However, streams and ditches that are missing from maps and databases are difficult to manage.

The main focus of this study was to develop a method to map channels combining deep learning and national Aerial Laser Scans (ALS). The performance of different topographical indices derived from the ALS data was evaluated, and two different Digital Elevation Model (DEM) resolutions were compared. Ditch channels and natural streams were manually digitized from ten regions across Sweden, summing up to 1923km of ditch channels and 248km of natural streams. The topographical indices used were: high-passing median filter, slope, sky-view factor and hillshade (with azimuths of 0°, 45°, 90° and 135°); while 0.5m and 1m were the DEM resolutions analysed. A U-net model was trained to segment images between ditches and stream channels: all pixels from each image were labelled in a way that those with the same class display similar attributes.

Results showed that ditches can be successfully mapped with this method and it can generally be applied anywhere since only local terrain indices are required. Additionally, when the natural streams are present in the dataset the model underperformed in predicting the location of ditches, while a higher resolution had the opposite effect. Streams were more challenging to map, and the model only indicated the channels, not whether or not they contained water. Further research will be required to combine hydrological modelling and deep learning.