SatelliteVision

Satellite image classification on xView with a fine-tuned Swin Transformer, imbalance-aware training, test-time augmentation and explicit evaluation by image identity.

Originally developed as MSc Computer Vision coursework and later refactored for reproducibility and stricter evaluation.

Results at a glance

Model	Split	Metric	Result
Swin-Base + TTA	validation	accuracy	83.20%

The historical run used 16,871 training images and 1,875 validation images across 13 classes.

The 83.20% is a validation result, not a held-out test score. The original checkpoint and per-image predictions were not preserved, so it is reported as historical evidence rather than as a result reproduced by the cleaned repository.

Highlights

• 13-class satellite crop classification.
• Full fine-tuning of microsoft/swin-base-patch4-window7-224.
• Class-weighted cross-entropy with MixUp/CutMix.
• Mixed-precision training and gradient accumulation.
• Four-view test-time augmentation (TTA).
• Prediction/label alignment by image ID and filename.
• Reproducible split manifests and lightweight integrity tests.

Pipeline

xView annotations
      ↓
ID-based train/validation split
      ↓
Swin-Base
      ↓
class weighting + MixUp/CutMix
      ↓
mixed precision + gradient accumulation
      ↓
validation + four-view TTA
      ↓
ID-aligned evaluation

The evaluator can report accuracy, fixed-13-class macro-F1, per-class precision/recall/F1, confusion matrix entries and misclassified examples for new prediction exports.

Repository structure

classification/     split and evaluation utilities
notebooks/          cleaned Swin training workflow
evidence/           historical outputs and split/provenance records
tests/              integrity and evaluation tests
VALIDATION.md       executed checks and known limits

Run

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python -m unittest discover -s tests -v
jupyter lab notebooks/swin_classification.ipynb

The original xView images are not redistributed. Point XVIEW_DATA to the local dataset before running training cells.

New prediction exports can be evaluated without retraining:

python -m classification.core \
  evidence/validation_split.csv \
  runs/swin/predictions.json

Reproducibility improvements

• persisted split identities;
• explicit image-ID/filename joins;
• rejection of missing or duplicated predictions;
• correct handling of the final gradient-accumulation group;
• structured metric/error exports;
• failure on unreadable images instead of silent substitution.

These changes improve the implementation but do not retroactively alter the historical 83.20% result.

Limitations

• Validation was used during development and is not an untouched final test set.
• The crop split does not explicitly control spatial or source-scene correlation.
• The original checkpoint and per-image historical predictions are unavailable.
• Training accuracy under MixUp/CutMix is not directly comparable with validation accuracy.
• This repository covers satellite classification, not object detection.

See VALIDATION.md for the exact checks executed during cleanup.