Automatic detection of ditches and natural streams from digital elevation models using deep learning

Policies focused on waterbody protection and restoration have been suggested to European Union member countries for some time, but to adopt these policies on a large scale the quality of small water channel maps needs considerable improvement. We developed methods to detect and classify small stream and ditch channels using airborne laser scanning and deep learning. The research questions covered the influence of the resolution of the digital elevation model on channel extraction, the efficacy of different terrain indices to identify channels, the potential advantages of combining indices, and the performance of a U-net model in mapping both ditches and stream channels. Models trained in finer resolutions were more accurate than models trained with coarser resolutions. No single terrain index consistently outperformed all others, but some combinations of indices had higher MCC values. Natural stream channels were not classified to the same extent as ditches. The model trained on the 0.5 m resolution had the most balanced performance using a combination of indices trained using the dataset with both types of channel separately. The deep learning model outperformed traditional mapping methods for ditches, increasing the recall from less than 10% to over 92%, while the recall for natural channels was around 71%. However, despite the successful detection of ditches, the models frequently misclassified streams as ditches. This poses a challenge, as natural channels are protected under land use management practices, while ditches are not. • The models can detect channels but struggle when classifying them into channel type. • Combining the Sky-view Factor and Slope had the best performance. • Combining topographic indices improved the model performance to classify streams.