Image-based bathymetry mapping for shallow waters

Abstract

Accurate bathymetric mapping is a key element for offshore activities, hydrological studies such as coastal engineering applications, sedimentary processes, hydrographic surveying as well as archaeological mapping and biological research. Through Structure from Motion (SfM) and Multi View Stereo (MVS) techniques, images can provide a low-cost alternative comparing to expensive LiDAR and acoustic surveys, offering as well, important visual information. Despite their relative low cost, their major drawback is that optical properties and illumination conditions of water severely affect overwater and underwater imagery and data acquisition process. However, when it comes to shallow waters, refraction seems to be the main factor affecting the geometry and the radiometry of the primary data and consequently of the results of overwater and underwater image-based 3D reconstruction methods. To that direction, the research carried out is concerned with the study and development of new methods for improving the performance and the accuracy of image-based mapping and 3D reconstructing the bottom in shallow waters for small- and large-scale surveys. The first method developed, addresses the systematic refraction errors on point clouds derived from SfM-MVS procedures in a generalized and accurate way. The developed method, based on Support Vector Machines, can accurately predict shallow bathymetric information from low altitude aerial image datasets over a calm water surface, supporting several coastal engineering applications in non-turbid waters and textured bottoms. The second method developed, is an image correction method which is built upon the state-of-the-art and firstly exploits the machine learning procedures that recover depth on the derived image-based dense point clouds (also a novel contribution of the thesis) and then corrects the refraction effect on the original imaging dataset. This way the operational SfM-MVS processing pipelines are ultimately executed on a refraction-free set of aerial imagery datasets resulting into highly accurate bathymetric maps and image-based products. This method also achieves a reduction on the noise of the sparse point clouds, which resulted from the SfM process and improves the accuracy and the quality of the produced orthoimages and textures. In the third method developed, an FCN classifier based on SegNet architecture is exploited in order to detect areas with caustics on the underwater imagery with very high accuracy, reliability and repeatability over different types of caustics, different types of seabed and luminosity conditions. The classifier is being trained using the first real world benchmark dataset on underwater caustics which is also a deliverable of the thesis. Having detected the rippling caustics on the initial imagery, a color transferring approach is performed, images are stereo-rectified and their respective disparity maps are generated. In the final step, the pixels classified as "caustics" are replaced by the corresponding pixels on the matched images that are classified as "non-caustics" using the disparity maps and the corrected stereo-rectified images are projected back onto the initial image model in order to facilitate further SfM and MVS processing and texturing. Results suggest an increment of 25% of the matched key-points between the images and more consistent 3D reconstruction, without any missing information. For all the developed methods, experimental results and validation over synthetic and real-world data are demonstrating their high potential, both in terms of bathymetric accuracy as well as texture and orthophoto quality. Also, frameworks were designed in a way to achieve high generalization over different platforms, cameras, flight heights, flight patterns and overlapping.