Multi-scale classification of photogrammetric point cloud data

Details

Supervisors

Defourny, Pierre ; Merle, Muru

Faculty

Faculté des bioingénieurs

Degree label

Master [120] en sciences agronomiques et industries du vivant, à finalité spécialisée: Geo-Information Science and Earth Observation for Environmental Modelling and Management

Abstract

Under the context of establishing a digital twin, the first step after data collection is to extract meaningful information from dense big data. This work provides a starting point by semantically labelling a photogrammetric point cloud dataset and seeks to find the optimal descriptors for extracting building information in a study area within Louvain-la-Neuve. A multi-scale point-wise classification approach was adopted where the point cloud was iteratively subsampled three times at three progressively larger grid sizes. At each subsample, 15 feature descriptors derived from height, reflectance, and eigenvalue-based geometry were computed using exclusively free and open-source software. The model with all features combined performed the best achieving an overall accuracy of 0.94, though with major confusion between roof and ground points. Although, geometric features performed the worst reaching a low of 0.59 overall accuracy, it was proven that the combination of geometric features with other color or height information led to more accurate classification of roofs and facades. Moreover, the reflectance information from photogrammetry proved to be highly valuable for vegetation extraction. In order to evade the confusion between roof and ground, the ground points were segmented using Cloth Simulation Filter (CSF) followed by the reconstruction of a gapless DEM. This approach showed great potential as since it enhanced the classification accuracy reaching 95% overall accuracy without the introduction of external datasets. However, further validation of roof points with a rooftop dataset from the Wallonian cartographic dataset (PICC) showed that a quarter of points classified as roof were needed to be removed to attain a 91% validation accuracy.