Deep Learning Approach using Patch-based Deep Belief Network for Road Extraction from Remote Sensing Imagery

In this paper, an automated Deep Learning Architecture (DLA) called patch-based Deep Belief Neural Networks (P-DBN) is designed, implemented and experimentally evaluated for extracting semantic maps of roads in Remote Sensing (RS) images. Representative features are extracted by unsupervised pre-training of DBN and supervised fine-tuning phase. A Logistic Regression (LR) is added to the end of feature learning system to constitute a P-DBN-LR architecture. This LR classifier is employed to fine-tune the whole pre-trained network in a supervised way and classifies the patches from RS images. The features extracted from the image patches are fed to the architecture as input and it produces the class labels as a probability matrix as either a positive sample (road) or a negative sample (non-road). A math morphology algorithm is used to improve P-DBN performance during post processing. Experiments are conducted on a dataset of 970 RS scene images of urban and suburban areas to demonstrate the performance of the proposed network architecture. The proposed deep model resulted in an Overall Accuracy (OA) of 95.57% and F1-score of 0.9588. When compared to other state-of-the-art deep learning-based models such as U-Net, Cascaded CNN, Roadtracer, FCN and Salient features-SVM, our proposed P-DBN-LR model outperforms by 3.22% (0.9550 vs. 0.9242), 4.56% (0.9550 vs. 0.9114), 3.30% (0.9550 vs. 0.9233), 5.54% (0.9550 vs. 0.9020) and 7.40% (0.9550 vs. 0.8843), for the dataset. Experimental results demonstrate the effective performance of the proposed method for extracting roads from a complex scene.