|M.Sc Student||Bauman Tal|
|Subject||Automatic Detection of Geospatial Changes based on|
Local Multi-Resolution Analysis of Digital
Elevation Models Produced by Drone
|Department||Department of Civil and Environmental Engineering||Supervisors||DR. Sagi Dalyot|
|DR. Ophir Regev Almog|
|Full Thesis text - in Hebrew|
High-performance and inexpensive Unmanned Aerial Vehicles (UAVs, Drones) are increasingly used for the acquisition of timely geospatial information (imagery) for the production of DEMs for geospatial change detection. DEMs produced from UAV imagery have very high resolution and very good internal (relative) accuracy. However, their absolute location accuracy is inferior to other mapping technologies. Therefore, existing change detection methods, which are based on the point-by-point comparison, will perform poorly when processing DEMs created from UAV imagery, since they are limited in reliably separating real physical changes from artifacts related to DEM inherent inaccuracy or errors. This research presents a novel methodology that overcomes these deficiencies, by implementing a hierarchical analysis and modeling process, in which a sequence of methods is used to automatically identify and match unique homological features, such as building corners or topographic maxima, in the various height models (DEMs). These provide geospatial anchors, which bring out local geospatial discrepancies between the models, used to "repair" (align) the models to the same geospatial reference system, at which local point-based change-detection is performed.
This study focuses on the development of a methodology based on multi-resolution analysis of DEMs, with the goal of retrieving optimal spatial transformation - or a set of such, which will enable accurate quantification of the changes that occurred, while maximizing the neutralization of contradictions that are not part of the changes. The first stage relies on finding and matching homological points between the various databases, which are characterized by uniform trends in their flat attribution. These enable the second phase of the extraction of spatially compatible regions locally, thereby ensuring the flexibility of spatial matching and maintaining continuity, allowing for more accurate and reliable detection of changes. In this way, it is possible for each local area in DEM A to adopt a spatial transformation to the homological zone in DEM B, in order to achieve maximum spatial compatibility between the databases, which will maintain continuity in the presentation of data.
Experimental results showed that when implementing the default point-by-point height differences process, 98.99% of the analyzed area was falsely classified as changed; whereas implementing the method developed in this research adequately detected all the changes occurred in the area with no false positives, correctly classifying 0.16% of the area as changed. Moreover, by applying the processes and algorithms developed in this research, very small changes in soil at depths of 10 cm and 40 cm in diameter are automatically detected, while at the same time eliminating the local error signal. These results validate the research hypothesis, suggesting new and robust processes and algorithms to handle automatic change detection in DEMs.