CNN vs ResNet18 vs ImprovedCNN
A land-use classification study comparing a baseline SimpleCNN, transfer learning with ResNet18 and an ImprovedCNN across agriculture, urban and other satellite-image classes.
Problem
A basic CNN can provide a useful baseline, but satellite-image categories can require stronger feature extraction and better generalization.
My Role
Designed the experiment workflow, prepared data splits, trained and compared models, and presented results through confusion matrices and model-size comparison.
Tools
- Python
- PyTorch
- torchvision
- SimpleCNN
- ResNet18 transfer learning
- ImprovedCNN
- Matplotlib
- scikit-learn
Dataset / Inputs
Land-use image classes: agriculture, urban and other. Train / validation / test splits with image preprocessing and evaluation metrics.
Results
ResNet18 achieved clearer class separation than the baseline, while ImprovedCNN reached 0.9311 accuracy and 0.9191 macro F1 in the project evaluation summary.
Challenges
Balancing model capacity with memory cost, improving difficult classes and showing results transparently through validation confusion matrices.
Demo & Screenshots
GitHub
Repository status: Private / available upon request. The public GitHub profile is linked below. For a walkthrough, code discussion, screenshots or a demo, contact me directly on WhatsApp.