End-to-End High-Quality DEM Generation and Ground Filtering From ALS Point Clouds

Ground filtering of airborne laser scanning (ALS) point clouds is crucial for producing high-quality digital elevation models (DEMs). Current learning-based methods commonly rely on point-wise classification, and minor errors may significantly degrade the quality of DEMs. Additionally, these methods are sensitive to point density variations. To address these challenges, this paper propose an end-to-end DEM generation method that directly produces DEMs from ALS point clouds. Specifically, the input point cloud is first converted into depth maps, where a Gaussian kernel is used to propagate smooth point elevations, enhancing robustness to varying point cloud densities. Then, a GAN-based model with elevation self-interpolation attention modules is employed to generate high-precision raster DEMs, eliminating the need for interpolation of DEMs from classified ground points. This approach not only produces more accurate terrain representation but also enables efficient ground filtering for ALS point clouds. Extensive experiments demonstrate that our method produces state-of-the-art performance across various scenarios with strong generalization capabilities, especially in complex landscapes.