Identification of morphologic and hydrodynamic features/patterns using image classification techniques

Evaluation of beach hydromorphological behaviour and its classification is highly complex. This complexity results from the interaction between wave climate and solid boundaries (beaches, groins, seawall, among others), presence of very dynamic events, nonlinearity of phenomena and interactions, different temporal scales (from second to hundred of years), and difficulty of obtain historical data (hydrodynamic, geomorphologic and topographic) reliable and continuous in time.
The development of new techniques on coastal studies, such as Geographic Information Technologies (GIT), using aerial photography and image processing techniques is an important issue on Coastal Engineering and Geo-Spatial Sciences research.
Several aerial photographs (between 1958 and 2003), using visual interpretation on a Geographic Information System (GIS) environment, were used on the identification of the coastal features/patterns in a selected area of the NW Portuguese coast, between Esmoriz and Mira beach (Pais-Barbosa et al., (2007) and Pais-Barbosa (2007)). This methodology allows for the identification and analyses of an important number of morphologic and hydrodynamic features/patterns in several aerial photographs datasets. However, this method presents two disadvantages: time consumption and the accuracy of the visual classification are not possible to evaluate.
Thus new analysis approaches with the aim of identify and analyse morphological and hydrodynamic features/patterns in aerial photographs, using image classification techniques (supervised and unsupervised classification), is under development. Therefore, the aim of this work, based on the image classification algorithms, is to identify and to analyse morphological and hydrodynamic features/patterns and to compare these results with the visual interpretation already performed.
Different supervised classification algorithms, such as parallelepiped, minimum distance and maximum likelihood were tested, in order to identify morphological and hydrodynamic features/patterns. Unsupervised classification algorithms (clustering) were also employed in order to determine the natural groupings or structures in the aerial photographs.
The performance and accuracy of the supervised and unsupervised classification algorithms were evaluated in order to validate the visual interpretation already performed. The supervised classification algorithm presents good performance, demonstrated by the results of the confusion matrix, Kappa coefficient and overall accuracy. The values of overall accuracy and Kappa statistic for the parallelepiped algorithm were 95.65% and 0.95661, respectively. The results for the maximum likelihood algorithm were similar: 95.85%, and 0.95840 for the overall accuracy and Kappa statistic, respectively. The minimum distance algorithm proved to be less accurate.
The non-supervised classification algorithms (K-means and ISODATA) allow for the identification of several classes, such as dry beach, wet beach and surf zone. The ISODATA algorithm is more flexible than the K-mean method. However, the user has to choose empirically more parameters. In the future will be investigated the unsupervised classification algorithms potentialities. The unsupervised classification algorithms have the advantage of needing less operator intervention.
The results of the supervised and unsupervised classification algorithms were compared with the visual identification and were in agreement with the visual interpretation. Overlapping the visual identification performed by Pais-Barbosa (2007) and Pais-Barbosa et al., (2007) and the parallelepiped algorithm (supervised classification) it can be observed a good agreement between the “classifications” tested. Moreover, the use of image classification algorithms reduced the time consumption and makes easier and more accurate the morphological and hydrodynamic features identification.