Comparative analysis of machine learning and point-based algorithms for detecting 3D changes in buildings over time using bi-temporal lidar data

Building Change Detection techniques are critical for monitoring building changes and deformations, construction progress tracking, structural deflections and disaster management. However, the performance of relevant algorithms on airborne light detection and ranging (lidar) data sets have not been comparatively evaluated, when such data sets are increasingly being used for construction purposes due to their capability of providing volumetric information of objects. This study aims to suggest appropriate building change detection algorithms based on a comparative evaluation of the performance of five selected algorithms including three pixel-based algorithms, Digital Surface Model differencing (DSMd), Support Vector Machine (SVM) and Maximum Likelihood (ML), and two point-based change detection algorithms, namely Cloud to Cloud (C2C) and Multiple Model to Model Cloud Comparison (M3C2). The algorithms were applied on two-point cloud samples from the same areas, and the results of pixel-based change detection algorithms indicate that the SVM algorithm could operate satisfactorily when noise is present in the data but could not reliably quantify the magnitudes of building height changes. The DSMd algorithm can derive the magnitudes of building height change, but it produces a high level of noise in the result and influences the change detection reliability. Therefore, an integration of DSMd and SVM was applied to determine the magnitudes of change and significantly reduce the noise in the results. Among point-based algorithms, M3C2 algorithm is able to show the magnitudes of building height changes and differentiate between new and demolished objects, while C2C can not fully satisfy the evaluation criteria. The authors recommend evaluation of these algorithms using additional temporal data sets and in various urban areas. Therefore, a generalization of the findings at this stage is premature.