CNNvsTransformer: Comparing CNN- and Transformer-Based Deep Learning Models for Semantic Segmentation of Remote Sensing Images

Master Thesis at the Institute for Geoinformatics, University of Münster, Germany.

Description

With the given code, results for the thesis mentioned above were created. The thesis is available as pdf file in this repository at \thesis.

In the following some basic steps to use the code are described but might need some adaptations, based on the utilized infrastructure.

Further documentation will be added soon.

How To Train the Models

1. Prepare data
   1. Download data, for example one of:
      1. ISPRS Benchmark on Semantic Labeling
      2. FloodNet
   2. Patchify data into appropriate sizes, e.g. $512\times 512$
      1. \helper\patchify.py might be useful
   3. Split data into train and test data in the following folder structure, with rgb folders containing the corresponding images to the groundtruth labels in the label folders:

|-- data
|   |-- rgb
|   |-- rgb_test
|   |-- label
|   |-- label_test

2. If you use the PALMA cluster, just adapt and run one of \PALMA\train_unet.sh or \PALMA\train_segformer.sh and you are done
3. Otherwise: Install requirements (see \PALMA\requirements.txt and modules in \PALMA\train_unet.sh)
4. Look at possible parameters in train.py and run the following line with respective adjustments:

python3 train.py --data_path /your/path/to/folder/data --name ./weights

5. Find your model in folder ./weights

By default a U-Net model will be trained for 20 epochs. Further default settings can be derived from the parameters of train.py.

Evaluation and Visualization

The evaluation and visualization of the models was done with help of the notebooks in the respective ./Notebooks directory.

• compare.ipynb: comparison of two models by visualizing predictions of both and calculating metrics on test data
• homogeneity.ipynb: calculation of clustering evaluation measures
• radarchart.ipynb: plot radar charts and bar charts for the calculated metrics
• count_classes.ipynb: count pixels per class and mean and standard deviation in an image dataset
• OUTDATED: Segformer_Run.ipynb, UNet_Run.ipynb, Segformer_visualize_attention.ipynb

Acknowledgements

A lot of the code was inspired by https://github.com/suryajayaraman/Semantic-Segmentation-using-Deep-Learning.