Satellite Image Classification with Hybrid CNN + Classical ML Models

---

Overview

This project investigates hybrid image classification pipelines for satellite remote sensing imagery. Rather than relying solely on end-to-end deep learning, we explore a two-stage approach: using a pretrained VGG16 CNN as a feature extractor, then benchmarking five classical machine learning classifiers on those extracted features.

The core question: can classical ML models match or complement a full deep network when given rich CNN-derived features?

The study is conducted on the RSI-CB256 dataset, a four-class satellite image benchmark, and provides a systematic comparison of task accuracy across all model configurations.

---

Dataset

RSI-CB256 — Remote Sensing Image Classification Benchmark

• 4 classes of satellite sensor imagery
• Images available in two resolutions: 224×224 and 64×64
• All images are standardized to 224×224 for consistency across experiments

---

---

Method

Pipeline Architecture

The system is built around a two-stage design:

Stage 1 — Feature Extraction (CNN) A VGG16 backbone pretrained on ImageNet is used as a fixed feature extractor via transfer learning. The output of the convolutional layers serves as a rich, high-dimensional feature representation of each satellite image.

Stage 2 — Classification The extracted CNN features are passed to one of two classification heads:

• Deep Network — A fully connected neural network trained end-to-end on the VGG16 features
• Classical ML Classifiers — The same CNN features are used to train and evaluate five classical models:

Classifier	Type
K-Nearest Neighbors (KNN)	Instance-based
Logistic Regression	Linear
Support Vector Machine (SVM)	Kernel-based
Random Forest (RF)	Ensemble / Bagging
AdaBoost	Ensemble / Boosting

This setup enables a controlled, apples-to-apples comparison: every classifier receives identical feature inputs, so differences in accuracy are attributable purely to the classifier, not the features.

---

Results

Deep Network — Training & Validation Performance

Training and validation loss and accuracy rate of deep network.

Full classifier comparison results and metrics are available in the notebook.

Key Takeaways

• Transfer learning is highly effective — VGG16 features pretrained on ImageNet generalize well to satellite imagery without any fine-tuning of the backbone
• Classical ML on CNN features is competitive — Several classical classifiers achieve strong accuracy when given rich deep features, demonstrating that end-to-end training is not always necessary
• SVM and Logistic Regression tend to perform best among the classical methods on linearly separable CNN feature spaces
• The hybrid pipeline offers a practical alternative in resource-constrained settings where full deep network training is costly

---

Topics

machine-learning deep-learning computer-vision image-classification transfer-learning vgg16 svm knn random-forest adaboost satellite-imagery remote-sensing jupyter-notebook

---

License

This project is licensed under the MIT License. See LICENSE for details.

---

Bursa Uludağ University · Computer Engineering Department