Semantic Segmentation on the Potsdam Dataset

This repository contains an implementation of semantic segmentation models trained on the 2D Semantic Labeling Potsdam dataset. The project explores two different deep learning approaches for pixel-wise classification of aerial imagery: a simple convolutional model and an encoder–decoder architecture.

The goal is to segment aerial images into six semantic classes:

• Impervious surface
• Building
• Tree
• Low vegetation
• Car
• Clutter/Background

---

Dataset

The dataset consists of GeoTIFF image tiles with a spatial resolution of 5 cm per pixel. Each tile has dimensions of 224 × 224 pixels and contains six bands:

• Red
• Green
• Blue
• Infrared (IR)
• Elevation
• Target label band (semantic classes)

For computational feasibility, 5,000 images were randomly sampled from the full dataset of 15,048 tiles.

The dataset was split into five folds:

Fold	Usage
1–3	Training
4	Validation
5	Test

To improve training efficiency, the dataset was stored in TFRecord format and loaded using the TensorFlow tf.data pipeline.

---

Models

Simple Convolutional Model

A baseline semantic segmentation model consisting of two convolutional layers followed by a softmax classifier.
This model uses RGB and IR bands as input and serves as a reference for performance comparison.

---

Encoder–Decoder Model

A more advanced encoder–decoder architecture was implemented using all five input bands:

• RGB
• Infrared
• Elevation

The encoder progressively downsamples the spatial resolution to extract high-level features, while the decoder reconstructs the segmentation map using transposed convolutions and skip connections.

---

Training

Both models were trained for 20 epochs using:

• Categorical Cross-Entropy Loss
• Adam Optimizer

Data augmentation was applied during training to improve generalization:

• Random horizontal flips
• Random vertical flips
• Random 90° rotations

The best model was selected based on validation loss using model checkpointing.

---

Results

The encoder–decoder architecture significantly outperformed the simple convolutional model.

Model	Test Loss	Test Accuracy
Simple CNN	0.924	63.4%
Encoder–Decoder	0.632	77.2%

---

Example predictions

Below are examples of model predictions compared to the ground truth.

RGB Image

Elevation Band

Ground Truth Segmentation

Model Prediction

---

Technologies used

• Python
• TensorFlow / Keras
• NumPy
• Matplotlib
• Rasterio

---

Authors

This project was implemented in Design of AI Systems in Chalmers, by:

Josue Valenzuela Perez
Ludwig Alexandersson