Road Detection of Remote Sensing Image Based on Convolutional Neural Network

Abstract

We propose an end-to-end framework for the road detection in satellite imagery with convolutional neural networks (CNNs). Firstly, we analyze the limitations of patch-based network and full convolution neural network and propose a new method. In our approach, CNNs are directly trained to produce classification maps out of the input images without losing much edge information. The method, called M-FCN, extends FCN by using a new activation function instead of the traditional activation function. We then address the issue of imperfect training data through a traditional approach: labeled images manually to form a new dataset. Finally, we show that such a network can be trained on remote sensing images with a composite loss function. At the same time, we validate the effect of label accuracy in dataset on the model. To ensure the accuracy of our method, we apply different methods to train in the same dataset. A series of experiments show that our networks consider a large amount of context to provide fine-grained classification maps, M-FCN outperforms SVM, patch-based network and full convolution network.