missingviews-study-eo's Introduction

Impact of missing views in multi-view learning models

A public repository of our work in missing views for Earth Observation (EO) applications.

Data

Preprocessed data can be accessed at: Link

Additional source

We use the following source for multi-view learning: https://github.com/fmenat/mvlearning

Training

To train a single-view learning model (e.g. Input-level fusion):

python train_singleview.py -s config/input.yaml

To train all the views individually with single-view learning (e.g. for single-view predictions or Ensemble-based fusion):

python train_singleview_pool.py -s config/pool.yaml

To train a multi-view learning model (e.g. Feature-level fusion, Decision-level fusion, Gated Fusion, Feature-level fusion with MultiLoss):

python train_multiview.py -s config/mv_feat.yaml

To train a multi-view learning model with CCA searching in case of missing views:

python train_multiview_cca.py -s config/mv_cca.yaml

Evaluation

To evaluate the model by its predictive quality:

python evaluate_predictions.py -s config/evaluation.yaml

To evaluate the model by its predictive robustness:

python evaluate_rob_pred.py -s config/evaluation.yaml

📜 Source

Public repository of our IGARSS 2023 paper.

🔓 Arxiv

🖊️ Citation

Mena, Francisco, et al. "Impact assessment of missing data in model predictions for Earth observation applications." IEEE International Geoscience and Remote Sensing Symposiums (IGARSS), 2024.

@inproceedings{mena2024igarss,
  title = {Impact assessment of missing data in model predictions for {Earth} observation applications},
  booktitle = {{IEEE International Geoscience} and {Remote Sensing Symposium} ({IGARSS})},
  author = {Mena, Francisco and Arenas, Diego and Charfuelan, Marcela and Nuske, Marlon and Dengel, Andreas},
  year = {2024},
  publisher = {{IEEE}},
}

Recommend Projects