An improved workflow for image- and laser-based virtual geological outcrop modelling
MetadataShow full item record
Photorealistic 3D models, representing an object’s surface geometry textured with conventional photography, are used for visualization, interpretation and spatial measurement in many disparate fields, such as cultural heritage, archaeology and throughout the earth sciences, including geology. Virtual models of geological outcrops allow for large quantities of geometric data, such as sizes of features, thicknesses of strata, or surface orientations to be extracted in relatively short time and in areas with difficult accessibility. However, standard analysis is limited to interpretation of the three standard spectral bands (red, green, blue; RGB) acquired in the visible spectrum by the conventional digital camera. Complementing the photorealistic 3D outcrop models with auxiliary spectral data, for example in the form of hyperspectral imagery, can provide domain experts with additional geochemical information, adding great potential to studies of mineralogy and lithology.
The existing workflows for creation of photorealistic outcrop models and integration with terrestrial panoramic hyperspectral data are complex and require specific knowledge from the field of geomatics. One such processing step is selection of images taking part in the texture mapping process. Although automated texture mapping measures are available, in highly redundant image sets they do not necessarily provide the best results when using all available photos. Therefore selection of the most suitable texture candidates is required to increase the realism of the textured models and the processing efficiency. Especially for large models of rugged terrain, represented by millions of triangles, manual selection of the best texture candidates can be challenging, because the user must account for occlusions and ensure that image overlap is sufficient to cover relevant model triangles.
The existing workflow for integration of hyperspectral and 3D data also requires specific skills in geomatics as homologous points between the two datasets need to be manually selected for registration. Finding such correspondences involves interpretation of data acquired with different sensors, in different parts of the electromagnetic spectrum, projections and resolutions. The need to complete such challenging data processing steps by users from outside the geomatics domain poses a serious obstacle to these methods becoming standardised across geological research and industry.
The research presented in this thesis addressed the two aforementioned limitations in the data processing workflows with an aim to make the method more accessible for users from outside of the geomatics domain. Firstly, a new interactive framework was developed, that provides analytical and graphical assistance in selection of an image subset for geometrically optimised texturing in photorealistic 3D models. Visualisation of spatial relationships between different components of the datasets was used to support the user’s decision in tasks requiring specific technical background. Novel texture quality measures were proposed and new automatic image sorting procedures, originating in computer vision and information theory, were implemented and tested. The image subsets provided by the automatic procedures were compared to manually selected sets and their suitability for 3D model texturing was assessed. Results indicated that the automatic sorting algorithms can be a valid alternative to manual methods. The resulting textured models were of comparable quality and completeness, and the time spent in time-consuming reprocessing was reduced. Anecdotal evidence indicated an increased user confidence in the final textured model quality and completeness.
Secondly, a method for semi-automatic registration of terrestrial hyperspectral imagery with laser and image data was developed. The proposed data integration procedure employed the Scale Invariant Feature Transform (SIFT) to automatically find homologous points between digital RGB images registered in the scanner coordinate system and short wave infrared cylindrical hyperspectral data. The need for large numbers of homologous points to be matched required optimisation of the SIFT operator, as well as a routine for eliminating false matches. The proposed method automatically provides the control points that are used for registering the hyperspectral imagery. The results obtained on two datasets with different characteristics indicated that the proposed method can be used as an alternative to manual data integration, saving time and minimizing user input during processing.
The increased automation of the workflows for creation of photorealistic outcrop models and integration with auxiliary image data, complemented with computer assistance to support users’ decision in the processing steps requiring background in geomatics, facilitate adoption of such techniques in wider community.