Full text loading...
Strong lateral variations in the subsurface break the assumptions behind Multichannel Analysis of Surface Waves (MASW). When using different spatial windows of receivers, the lateral variations are smeared out through dispersion spectra. We attempt to remove the effects of lateral variations by considering this a deblurring problem. In analogy with deblurring of visual images, we train a convolutional neural network to perform the task. The training data is formed by synthetic data with a sharp boundary either around the centre or at a random location in the model. When the location of the boundary is roughly known, the results improve clearly. However, when the boundary can be anywhere, the trained model is not able to fully remove the effects of lateral variations.