Roof type classification with innovative machine learning approaches.

Area of Science:

• Computer Science
• Remote Sensing
• Artificial Intelligence

Background:

• Convolutional neural networks (CNNs) are widely used for image classification but require extensive training data.
• The scarcity of labeled roof images poses a significant challenge for CNN-based roof type classification.
• One-shot learning mimics human learning by enabling classification with very few examples per category.

Purpose of the Study:

• To investigate the effectiveness of one-shot learning for roof type classification using Siamese neural networks.
• To compare the performance of a Siamese neural network model against a traditional CNN model for this task.
• To address the data scarcity issue in roof type classification.

Main Methods:

• Utilized a Siamese neural network architecture for one-shot learning.
• Generated artificial images for training due to the difficulty in obtaining real roof data.
• Employed a dataset of real roof images (flat, gable, hip) for testing.
• Trained both a CNN-based model and a Siamese neural network model on the same dataset.

Main Results:

• The Siamese neural network model, trained on artificially generated images, achieved an average classification success rate of 66% on real roof images.
• This demonstrates the potential of one-shot learning in scenarios with limited training data.
• Performance comparison between the Siamese network and the CNN model was conducted.

Conclusions:

• One-shot learning, particularly with Siamese neural networks, offers a viable solution for roof type classification when training data is scarce.
• The 66% accuracy achieved highlights the practical applicability of this method in remote sensing and urban planning.
• Further research can explore more sophisticated data augmentation and network architectures to improve performance.