Aerial Image Segmentation & Land-Cover Classification

A geospatial machine learning project for segmenting and classifying high-resolution aerial imagery using RGB, infrared (IR), and Digital Surface Model (DSM) data.

The project compares traditional image segmentation and machine learning methods with a deep learning approach for semantic land-cover classification.

Project Overview

The objective was to extract meaningful spatial objects and classify land-cover categories from high-resolution aerial imagery.

The workflow covers image preprocessing, unsupervised segmentation, feature extraction, Random Forest classification, and CNN-based semantic segmentation.

Main Workflow

• Preparation and exploration of RGB, IR, and DSM imagery
• Chessboard segmentation
• K-Means clustering
• SLIC superpixel segmentation
• Feature extraction from image segments
• Random Forest classification
• CNN-based semantic segmentation
• Evaluation and visualization of classification results

Dataset

The project uses high-resolution aerial imagery from the ISPRS Potsdam dataset, including:

• RGB imagery
• Infrared information
• Digital Surface Model (DSM)
• Reference land-cover labels

The combination of spectral and elevation information provides complementary information for distinguishing urban objects and land-cover classes.

Input Data

Image Segmentation

Several segmentation approaches were explored before semantic classification.

K-Means

K-Means clustering was used as an unsupervised approach to group pixels based on similarity in feature space.

SLIC Superpixels

SLIC was used to divide the imagery into compact superpixels that better follow local object boundaries. These segments can then be used for object-based feature extraction and classification.

Random Forest Classification

A Random Forest classifier was used for supervised land-cover classification. Features derived from the imagery and segmented regions were used to predict semantic classes.

Evaluation

The classification results were compared with the available reference labels to assess model performance.

CNN-Based Semantic Segmentation

A convolutional neural network was also explored for pixel-level semantic segmentation. This extends the project from traditional machine learning to a deep learning approach for learning spatial and spectral patterns directly from the imagery.

Tools & Technologies

• Python
• Jupyter Notebook
• NumPy
• Matplotlib
• scikit-learn
• scikit-image
• PyTorch
• Random Forest
• K-Means
• SLIC Superpixels
• Convolutional Neural Networks (CNN)
• RGB / IR imagery
• Digital Surface Models (DSM)

Repository Structure

Aerial-Image-Segmentation/
├── README.md
├── notebooks/
│   ├── 01-image-segmentation.ipynb
│   ├── 02-random-forest-classification.ipynb
│   └── 03-cnn-semantic-segmentation.ipynb
└── images/
    ├── 01-input-data-overview.png
    ├── 02-slic-superpixels.png
    ├── 03-random-forest-classification.png
    ├── 04-random-forest-evaluation.png
    └── 05-cnn-semantic-segmentation.png

Key Skills Demonstrated

Geospatial machine learning · Pattern recognition · Image segmentation · Feature extraction · Supervised and unsupervised learning · Random Forest · Deep learning · CNNs · Semantic segmentation · Remote sensing data analysis