SatDetection

Multi-class semantic segmentation of aerial imagery using U-Net + ResNet-50.

Course: CSCI 4366/6366 — Neural Networks and Deep Learning, Spring 2026
Team: Rufai Yakubu, Aaron Tyler
Affiliation: The George Washington University

---

Overview

Pixel-level land cover classification from high-resolution aerial imagery using the LandCover.ai dataset.

Classes:

Class	Mask Value
Background	0
Building	1
Woodland	2
Water	3
Road	4

---

Dataset

• Source: LandCover.ai — aerial imagery from Poland, 25 cm/pixel resolution
• Size: 41 GeoTIFF images, auto-downloaded via torchgeo
• Tiling: Each 512×512 image is split into four 256×256 tiles → stored under data/tiles/

---

Project Structure

SatDetection/
├── main.py              # Entry point — tiling, training, evaluation
├── SegDataset.py        # Dataset class (loads tiled images and masks)
├── SegModel.py          # Training loop, validation, testing, visualization
├── UNetResNet50.py      # Custom U-Net with ResNet-50 encoder
├── img_helpers.py       # Tiling, preview, and plotting utilities
├── scr.py               # Scripts / scratch
├── test.py              # Standalone test script
├── data/
│   └── tiles/
│       ├── images/      # Tiled RGB patches (.png)
│       └── masks/       # Tiled label masks (.png)
├── out/                 # Output visualizations
├── models/              # Saved model checkpoints (.pt)
└── requirements.txt

---

Setup

pip install -r requirements.txt

---

Usage

python main.py

On first run this will:

1. Download the LandCover.ai dataset via torchgeo
2. Tile images into 256×256 patches under data/tiles/
3. Train the model and save the best checkpoint to models/
4. Run evaluation and display a sample prediction

Key flags in main.py:

Flag	Default	Description
load_model	False	Load a saved model instead of training
custom_unet	False	Use custom UNetResNet50 instead of smp.Unet
weights	'imagenet'	Encoder weights ('imagenet' or None)
model_name	'best_model_untrained'	Name used for saving/loading

---

Model

Two model options are supported:

• smp.Unet (default) — U-Net with ResNet-50 encoder from segmentation-models-pytorch
• UNetResNet50 — custom U-Net with ResNet-50 backbone built with torchvision

Both support optional ImageNet pretrained weights on the encoder.

Architecture summary:

• Encoder: ResNet-50 (5 stages)
• Decoder: transposed convolutions with skip connections
• Head: 1×1 conv → 5-class logits
• Loss: Dice loss (multiclass)
• Optimizer: Adam (lr=3e-4)
• Metric: Mean IoU

---

Results

Best checkpoint saved to models/ based on validation mIoU. Sample predictions are visualized with per-class color overlays:

Class	Color
Background	Black
Building	Red
Woodland	Green
Water	Blue
Road	Gray