Skip to content
Snippets Groups Projects

Advanced Deep Learning Architectures for Accurate Detection of Subsurface Tile Drainage Pipes from Remote Sensing Images

This git page contains the code used in the paper "Advanced Deep Learning Architectures for Accurate Detection of Subsurface Tile Drainage Pipes from Remote Sensing Images", presented at the SPIE Image and Signal Processing conference in Berlin September 2022. This README only contains an overview of the repository. More detailed explanations can be found in each subfolder.

If you use the code from this repository in your research, please cite the following paper:

@article{dlpipe2022,
  title={Advanced Deep Learning Architectures for Accurate Detection of Subsurface Tile Drainage Pipes from Remote Sensing Images},
  author={Tom Lukas Breitkopf and Leonard Hackel and Mahdyar Ravanbakhsh and Anne-Karin Cooke and Sandra Willkommen and Stefan Broda and and Begüm Demir},
  booktitle={SPIE Sensors + Imaging 2022},
  year={2022}
} 

Drainage Pictogram

Abstract

Subsurface tile drainage pipes provide agronomic, economic and environmental benefits. By lowering the water table of wet soils, they improve the aeration of plant roots and ultimately increase the productivity of farmland. They do however also provide an entryway of agrochemicals into subsurface water bodies and increase nutrition loss of soils. For maintenance and infrastructural development, accurate maps of tile drainage pipe locations and drained agricultural land are needed. However, for multiple reasons, these maps are often outdated or not present. To overcome these restrictions different remote sensing image processing techniques have been applied over the years with varying degrees of success. Recent developments in deep learning (DL) techniques improve upon the conventional techniques with machine learning segmentation models. In this study, we introduce two DL-based models: i) improved U-Net architecture; and ii) Visual Transformer in the framework of the tile drainage pipe detection. Our research shows that both models can improve the results in the dice coefficient and the Intersection over Union (IoU) over the baseline resulting in a dice coefficient of 0.68 and 0.70 for the improved U-Net and the Visual Transformer and an IoU of 0.80 and 0.81 respectively.

Structure of the repository

For either of the models, there is a separate folder in the repository, each containing detailed information on reproducing the results as shown in the report as well as the trained weights.

  • Improved UNet: Contains the code for the improved U-Net. This is one of the models newly adapted to the task of drainage pipe detection.
  • TransUNet: Contains the code for the TransUNet. This is the other model newly adapted to the task of drainage pipe detection.
  • UNet-Tile: Contains the code for the UNet-Tile model. This is the baseline model used to compare the newly introduced models to. Please note, that the code in this folder is mainly the baseline code as provided by the authors of the original paper and only the notebook was created to create a comparable evaluation for all models. The original code is introduced in this project.

Data

The data used for this project is stored in a Google Drive, as provided in the UNet-Tile project.

Weights

The weights of pre-trained models are stored within each individual folder.

This work is funded by the European Research Council (ERC) through the ERC-2017-STG BigEarth Project under Grant 759764 and by the German Ministry for Education and Research as BIFOLD - Berlin Institute for the Foundations of Learning and Data (01IS18025A).