ML/FOS detection system for drought induced cracks on dikes

Manmade soil-based flood defences are conceived and designed for containing water outside flood prone areas during extreme wet events. However, their reliability is reduced in time due to extreme drought events due to formation of drought induced cracks. While monitoring and maintenance are robustly done by visual expert inspections, they are not sufficiently efficient given the length and heterogeneity of soil retaining defences such as dikes and dams. In the present study, we explore the feasibility of a machine learning based fiber optic sensing system for detecting cracks over heterogeneous soil dikes. The system consists in generating detailed output training datasets from thermal imagery of both cracked and healthy dike soil and combining it with FOS thermal signal for the same exact locations. These datasets were collected during the summer of 2021 in the dike lab testing facility (Flood Proof Holland) from TU Delft in the Netherlands. If successful, the system will allow to use the FOS distributed signal in long dike stretches to detect un-observed cracked locations that present similar FOS signal in space and time. The preliminary results show great potential but it still remains to test it in significantly larger dike stretches and during dryer periods.