Extraction of Digital Terrain Models from Airborne Laser Scanning Data based on Transfer-Learning

Abstract

With the rapid urbanization, timely and comprehensive urban thematic and topographic information is highly needed. Digital Terrain Models (DTMs), as one of unique urban topographic information, directly affect subsequent urban applications such as smart cities, urban microclimate studies, emergency and disaster management. Therefore, both the accuracy and resolution of DTMs define the quality of consequent tasks. Current workflows for DTM extraction vary in accuracy and resolution due to the complexity of terrain and off-terrain objects. Traditional filters, which rely on certain assumptions of surface morphology, insufficiently generalize complex terrain. Recent development in semantic labeling of point clouds has shed light on this problem. Under the semantic labeling context, DTM extraction can be viewed as a binary classification task. This study aims at developing a workflow for automated point-wise DTM extraction from Airborne Laser Scanning (ALS) point clouds using a transfer-learning approach on ResNet. The workflow consists of three parts: feature image generation, transfer learning using ResNet, and accuracy assessment. First, each point is transformed into a feature image based on its elevation differences with neighbouring points. Then, the feature images are classified into ground and non-ground using ResNet models. The ground points are extracted by remapping each feature image to its corresponding points. Lastly, the proposed workflow is compared with two traditional filters, namely the Progressive Morphological Filter (PMF) and the Progress TIN Densification (PTD). Results show that the proposed workflow establishes an advantageous accuracy of DTM extraction, which yields only 0.522% Type I error, 4.84% Type II error and 2.43% total error. In comparison, Type I, Type II and total error for PMF are 7.82%, 11.6%, and 9.48%, for PTD are 1.55%, 5.37%, and 3.22%, respectively. The root mean squared error of interpolated DTM of 1 m resolution is only 7.3 cm. Moreover, the use of pre-trained weights largely accelerated the training process and enabled the network to reach unprecedented accuracy even on a small amount of training set. Qualitative analysis is further conducted to investigate the reliability and limitations of the proposed workflow.