xView Satellite Imagery Classification

(Deep Learning Course Project – Universidad Politécnica de Madrid, Master in Digital Innovation – EIT Digital)

Authors:

• Ádám Földvári
• Joseph Tartivel
• Máté Lukács

Instructor: Roberto Valle

📖 Overview

This repository contains the final project for the Deep Learning course at UPM. The objective was to classify objects in high-resolution satellite imagery using progressively advanced deep learning techniques:

• Feedforward Neural Networks (FFNN)
• Regularized FFNNs
• Convolutional Neural Networks (CNNs)
• Transfer Learning with ResNet50

📊 Dataset

• Source: xView Dataset
• Type: High-resolution satellite images (0.3m GSD, WorldView-3)
• Split: 761 training images, 85 test images
• Processed: 21,377 training objects and 2,635 test objects (cropped & resized to 224×224)
• Classes: 12 categories (e.g., building, small car, cargo plane, helicopter)

🧪 Evaluation Platform

Final testing and evaluation were conducted via a private competition on Codabench. Submissions were provided in the required JSON format and benchmarked against a hidden test set.

📂 Project Structure

.
├── ffnn.ipynb # Feedforward Neural Network experiments
├── reg.ipynb # Regularization strategies for FFNNs
├── cnn.ipynb # Custom Convolutional Neural Networks
├── tl.ipynb # Transfer Learning with ResNet50
├── DeepLearning_JT_AF_ML_finalReport.pdf # Full technical report
└── README.md # Project overview and methodology

Each notebook corresponds to a development phase, with models iteratively refined at each stage.

🔎 Methodology & Results

Phase 1 – Feedforward Neural Networks (FFNN)

• Compared shallow vs. deep architectures using flattened image inputs.
• Observed overfitting in deeper models due to loss of spatial structure.
• Best test accuracy: 45.2%

Phase 2 – Regularized FFNNs

• Applied batch normalization, dropout, and extended training.
• Improved generalization significantly.
• Best test accuracy: 55.18%

Phase 3 – Convolutional Neural Networks (CNNs)

• Developed and refined five CNN architectures.
• Integrated data augmentation, L2 regularization, batch normalization, dropout, and custom LR schedules.
• Best test accuracy: 76.36%

Phase 4 – Transfer Learning (ResNet50)

• Employed a two-stage strategy:
  • Feature extraction with frozen layers.
  • Selective fine-tuning of top layers with lower LR.
• Achieved highest overall performance with reduced training time.
• Best test accuracy: 77.87%

📈 Results Summary

Model	Test Accuracy	Precision	Recall
FFNN (Simple)	45.2%	30.31%	33.66%
FFNN + Regularization	55.18%	44.95%	55.50%
Custom CNN	76.36%	74.00%	74.53%
Transfer Learning (ResNet50)	77.87%	67.15%	77.47%

💡 Key Insights

• Spatially-aware architectures (CNNs, ResNet50) are critical for image classification tasks.
• Regularization substantially improves generalization for non-convolutional models.
• Class imbalance remains a challenge, especially for minority categories (e.g., helicopters).
• Transfer learning offered the best trade-off between accuracy, recall, and development time.

⚙️ Infrastructure & Evaluation

• Platform: Kaggle with P100 GPU acceleration
• Evaluation: Codabench private competition with hidden test set
• Submission format: JSON files for leaderboard evaluation

📄 Documentation

For complete methodology, experiments, and analysis, see the Final Report.