Optimized Deep Network for Precise Digital Terrain Model Extraction from Light Detection and Ranging Data

Classifying and filtering non-ground objects from the point cloud data are among the major challenges to the development of a digital terrain model (DTM). This paper proposes a hierarchical deep network to filter and classify non-ground objects from the point cloud data. The proposed network is mainly based on a deep encoder-decoder network with effective convolutional connections for extracting and fusing the features of shallow and deep layers to detect the objects better. In the proposed encoder-decoder network, a feature extraction block was designed to extract various features at different levels. The network also adopts a global-local feature fusion strategy. The proposed hierarchical deep network is based on the extraction and gradual fusion of features on different scales to extract objects of various dimensions and densities in urban areas with different topographical conditions. Evaluation of the proposed deep network in US and ISPRS datasets indicated its high accuracy of object detection in complicated and dense areas. In the US dataset, our model reduced total error by 0.055 and increased kappa by 20.18% on average compared to the second-best method. According to the results of a performance analysis through the ISPRS data, the proposed deep network outperformed the other methods and reduced the total error by 0.130 compared to the existing methods.