Automatic lake’s bottom detection from GPR B-scans using convolutional neural networks

BARBOSA HETHERINGTON, ANDRES; VILLAROSA, GUSTAVO; BEIGT, DÉBORA

Bariloche

Congreso; IAL IPA 2022; 2022

International Paleolimnology Association - International Association of Limnogeology

Ground Penetrating Radar (GPR) has proven to be an excellent technique to profilinglake’s underwater environments. Either when a model inversion is performed or whengeomorphology studies are done, it is necessary to pick the lake’s bed reflection. This is atime-consuming task that usually is done manually picking trace by trace the reflection. In this work a convolutional neural network is proposed to automate this task. This method consistsin three steps: The first one is to choose a sample size to split each trace and every sampleis labeled as positive (bottom’s lake) or negative (water column and other reflections). Thesecond is to design a convolutional neural network adequate for the problem and train itwith the labeled data in a supervised manner. Lastly a post-processing step is performed toimprove the results and ensure each trace only gets one prediction. Also, predictions withdifferent types of pre-processing are compared to the raw data. The network was trainedwith data from the Frías and Nahuel Huapi (Brazo Blest) lakes and tested in the Espejo lakegetting preliminary results as good as 95% accuracy on the test data.