Classification of Seabed Types by OBIA With Recursive Feature Elimination Corrected by Expert Interpretation

The characteristics of seabed types are of great importance for understanding the marine environment and the changes taking place there, spatial planning, seabed investments, and natural resources extraction. Currently, machine learning methods and expert interpretations are often used in the classification of seabed types. Despite the strong development of automated methods in natural science research, which significantly accelerate the analysis of large areas, their results are not fully consistent with reality. The validation of semi-automatic classification results also remains an important issue. Expert analysis, although time-consuming and sometimes subjective, remains the most accurate tool for identifying bottom classes, but in practice it is very difficult to perform it over large areas. In order to reduce these problems, the authors present a data processing scheme starting with data registration, through machine learning, determining probability and accuracy of prediction, determining hybrid classes (combining information about possible types of bottom) and finally expert interpretation using the results of semi-automatic classification. This minimizes the time spent by experts and reduces subjectivity, identifies areas with low probability where expert interpretation is very valuable, and improves the semi-automatic classification result. The work was performed using point information on seabed sediment types and bathymetric and backscatter data recorded with multibeam echosounder (MBES), on the basis of which second-order parameters were prepared. The CatBoost training algorithm and Recursive Feature Elimination with Cross-Validation (RFECV) were used. The average accuracy score from 50 repetitions of the predictive algorithm was 75%. Using the results of semi-automatic classification, developed hybrid classes, probabilities, and originally recorded data, the expert prepared a map of bottom types, correcting errors in semi-automatic classification. The presented results are valuable due to the described semi-automatic classification procedure applied to marine data, as well as the use of hybrid classes and the indication of the importance of combining the capabilities of machine learning methods with expert knowledge.