Extraction of rocks and boulders on natural terrain using semantic segmentation

The geological mapping for the rock outcrop and boulders distribution on countryside is necessary for support hazard and risk management. Different approaches were proposed to detect and identify the rock outcrop and boulders based on remote sensing technology in literature. Given the latest development in technology, using traditional remote sensing methods to detect the rocks and boulders in VHR aerial imagery is still challenging. The complicated background noise prohibits the precision and accuracy of identifying the targets as well as their boundaries detection. Currently, deep neural networks (DNNs) have been demonstrating its outstanding performance on feature extraction which has been widely used in the computer vision. The semantic image segmentation is a pixel-level classification model, which is appropriate to the rock and boulder detection due to the model itself can segment very dense and detailed features over VHR imagery. In this paper, an optimized FCN-DenseNet was proposed to detect the rock outcrops and boulders in Hong Kong. Since the sizes of the rocks and boulders are variated, the neural network should be employed with multi-scale pooling to extract low/high-level features. As the contexts of the rocks and background terrain are similar, the pyramid dense blocks were employed to enhance the capability of self-feature-extraction. Our proposed boulder-net was evaluated over the VHR imagery in Hong Kong. The results show that the rocks and boulders are identified and classified and the boundaries of the rocks are accurately segmented.