This repository holds all the necessary code to run the very-same experiments described in the paper "Creating Classifier Ensembles through Meta-heuristic Algorithms for Aerial Scene Classification".

If you use our work to fulfill any of your needs, please cite us:

@inproceedings{Ferreira:21,
  author={Ferreira, Álvaro R. and de Rosa, Gustavo H. and Papa, João P. and Carneiro, Gustavo and Faria, Fabio A.},
  booktitle={2020 25th International Conference on Pattern Recognition (ICPR)}, 
  title={Creating Classifier Ensembles through Meta-heuristic Algorithms for Aerial Scene Classification}, 
  year={2021},
  volume={},
  number={},
  pages={415-422},
  doi={10.1109/ICPR48806.2021.9412938}
}

• data/
  • RSDataset: Folder containing the RSDataset data;
  • RSSCN7: Folder containing the RSSCN7 data;
  • UCMerced_LandUse: Folder containing the UCMerced_LandUse data;
• models/
  • ensemble.py: Ensemble-based methods, such as weight-based and majority voting;
• utils/
  • constants.py: Constants definitions;
  • dictionary.py: Creates a dictionary of classes and labels;
  • load.py: Loads the dataset according to desired format;
  • metrics.py: Provides several metrics calculations;
  • mh.py: Wraps the meta-heuristic classes;
  • wrapper.py: Wraps the optimization tasks;

You may just install the pre-needed requirements under your most preferred Python 3+ environment (raw, conda, virtualenv, whatever):

pip3 install -r requirements.txt

Our first ensemble-based baseline is to use the Majority Voting in order to create a count-based ensemble. With that in mind, just run the following script with the input arguments:

python3 majority_voting.py -h

Note that it will output a helper file in order to assist in choosing the correct arguments for the script.

After defining the Majority Voting baselines, now we can proceed and try to find the most suitable weights for the ensemble (one can use a weight-based or a boolean-based approach) through a meta-heuristic optimization process. Just choose the following scripts and invoke their helper:

python3 ensemble_learning.py -h

and

python3 ensemble_learning_with_gp.py -h

and

python3 ensemble_learning_with_umda.py -h

Note that Genetic Programming- and Univariate Marginal Distribution Algorithm-based optimization are included in different scripts due to their particular structure defined in the Opytimizer library.

Finally, after concluding the optimization step over the validation sets, it is now possible to load back the best weights found during the optimization procedure and apply them into a weight-based ensemble over the testing set. Run the following script in order to fulfill that purpose:

python3 process_optimization_history.py -h

Note that the optimization process will always output a .pkl file, while the other scripts will output a *.txt file.

Instead of invoking every script to conduct the experiments, it is also possible to use the provided shell script, as follows:

./pipeline.sh

Such a script will conduct every step needed to accomplish the experimentation used throughout this paper. Furthermore, one can change any input argument that is defined in the script.

As an optional procedure, one can also calculate some diversity metrics between classifiers. Please use the following script in order to accomplish such an approach:

python3 diversity_metrics.py -h

Note that this script will also calculate both classifier accuracies over the desired dataset and fold.

We know that we do our best, but it is inevitable to acknowledge that we make mistakes. If you ever need to report a bug, report a problem, talk to us, please do so! We will be available at our bests at this repository or aalvin10@gmail.com and ffaria@unifesp.br.