Roadscape AutoEncoder: Redefining Road Detection with Semantic Segmentation

Road detection from high-resolution satellite images (HRSI) is critical for various applications such as urban planning and emergency response, yet manual approaches are labour-intensive and prone to errors. Existing methods employ supervised and unsupervised learning techniques, but challenges persist due to image complexity and noise. In this study, we propose a novel approach that combines semantic segmentation with autoencoder neural networks to generate precise road segmentation masks. Leveraging the Deep Globe Road Extraction Dataset and an independent dataset from Google Maps satellite imagery, we validate our method's effectiveness. Our CNN-based autoencoder architecture extracts hierarchical features using convolutional layers and optimizes model parameters with optimization algorithms. Through extensive training, the autoencoder learns to produce accurate road segmentation masks by minimizing reconstruction loss. Experimental results demonstrate remarkable accuracies of 95.39% in training and 95.17% in validation, showcasing the model's robustness. Evaluation on diverse scenarios from the “Satellite-Google maps-Masks” dataset further validates the practical utility of our method in urban development and road infrastructure research.