| --- |
| license: mit |
| language: |
| - en |
| pipeline_tag: image-classification |
| library_name: keras |
| metrics: |
| - accuracy |
| - f1 |
| tags: |
| - computer-vision |
| - satellite-imagery |
| - remote-sensing |
| - xview |
| - efficientnet |
| - ensemble |
| - transfer-learning |
| - tensorflow |
| --- |
| |
| # EfficientNetB0 Ensemble for Satellite Image Classification |
|
|
| An ensemble of three ImageNet-pretrained EfficientNetB0 classifiers for satellite image classification on a 13-class subset of xView, developed as a Deep Learning course project at Universidad Politécnica de Madrid (UPM). |
|
|
| ## Architecture |
|
|
| - **Task:** single-label classification of satellite image crops into 13 classes. |
| - **Input:** RGB images resized to 224 × 224 pixels. |
| - **Backbone:** EfficientNetB0 with ImageNet-pretrained weights, frozen during training. |
| - **Classification head:** custom dense layers trained for the target classes. |
| - **Ensemble:** three classifiers combined by averaging their predicted class probabilities (soft voting). |
| - **Framework:** TensorFlow / Keras. |
|
|
| This classifier operates on image crops, it does not locate objects in full satellite scenes. |
|
|
| ## Results |
|
|
| | Evaluation split | Metric | Result | |
| | --- | --- | --- | |
| | Validation | Accuracy | 83.96% | |
| | Validation | Macro F1 | 86.43% | |
| | Test (course Codabench evaluation) | Accuracy | 82.66% | |
|
|
| Validation results are recorded in the original notebook, the test score is reported in the project report. These figures describe the course's 13-class classification setup, not the full xView object detection benchmark. |
|
|
| ## Code and reproducibility |
|
|
| The relevant training notebook is `CNN Transfer Learning Best model.ipynb`. The accompanying project report is `Report_ImageRecognitionAndObjectDetectiononthexViewSatelliteDataset.pdf`. |
|
|
| To obtain a usable model, the classifiers must be trained again using the required dataset and preprocessing pipeline. Any new training run must be evaluated separately; the original scores are not guaranteed to be reproduced. |
|
|
| ## Intended use and limitations |
|
|
| Intended for educational use and review of the original methodology. Performance on other sensors, geographic regions, or image distributions has not been established. No production deployment or operational validation is claimed. |
|
|
| ## Authors |
|
|
| Melen Laclais, Léo Lamy, and Adrián García-Pozuelo Fornieles. |
|
|
| ## License and dataset attribution |
|
|
| The MIT license designation applies to original project code only. Third-party code and course materials retain their respective terms. |
|
|
| xView imagery and annotations are distributed under **CC BY-NC-SA 4.0** and are not covered by the MIT license. Their attribution, noncommercial-use, and share-alike requirements remain applicable, including to dataset images reproduced in notebooks or the report. |
|
|
| - [xView dataset](https://xviewdataset.org/) |
| - [Official xView dataset license terms](https://challenge.xviewdataset.org/rules) |
| - [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/) |
|
|
| Dataset reference: Lam et al., *xView: Objects in Context in Overhead Imagery* (2018). |
|
|
