Deep Learning-Based Segmentation of Land Drainage Systems from High-Resolution Aerial Imagery

Approximately 30% of the agricultural land in the Czech Republic (around 1 million hectares) has been drained, which significantly affects the water regime and water availability in the landscape. The largest expansion of agricultural drainage systems occurred during the communist and socialist era, particularly between the 1950s and 1980s. Old paper project documentation for these systems has often not been preserved, archival records are highly fragmented, and many existing plans do not correspond to the actual implementation.

Accurate mapping of drainage systems is essential for understanding detailed hydrological processes in the landscape, as well as for designing measures to mitigate their negative impacts. Automatic detection of drainage systems represents an important step toward comprehensive mapping of functional drainage structures. Traditional approaches based on manual interpretation of aerial imagery are time-consuming and practically infeasible for large areas. Therefore, the presented project proposes and tests a method using convolutional neural networks for the segmentation of drainage lines from high-resolution aerial imagery. The aim is to assess the potential of up-to-date machine learning techniques for automated extraction of drainage systems in landscapes with varying vegetation cover.

A critical component of the workflow is the preparation of training data, including manual annotation of drainage lines, creation of mask layers, and data augmentation to enhance model generalization. Preliminary results will be presented, including segmentation examples and discussion of key limitations, such as sensitivity to vegetation cover.

Segmentation is implemented using the U-Net architecture, widely applied for pixel-level classification tasks in geosciences. The encoder is based on ResNet34, enabling hierarchical feature extraction and improving robustness to texture and illumination variability in aerial imagery. The implementation was carried out in PyTorch using the Segmentation Models PyTorch library. Skip connections between corresponding levels ensure preservation of spatial details and accurate localization of linear structures typical of drainage systems.

Results indicate that deep neural networks significantly accelerate and improve the accuracy of drainage feature identification, opening new possibilities for various landscape analyses, incl. hydrology, agricultural management, landscape planning, nature protection etc. Future work will focus on expanding the training dataset, optimizing hyperparameters, and validating the model on a large set of aerial images.