Lineamentos topográficos em modelos digitais de elevação: Uma comparação entre extração manual e computacional usando redes neurais convolucionais com arquitetura U-Net

Abstract

Topographic lineaments seen on digital terrain models (DTM) provide good insights into geological features. To extract useful information from these data interpretation is a crucial step and is traditionally done manually by a geoscientist in a repetitive and time-consuming process, prone to human bias and repetitive stress injury. Recent developments in the application of machine learning techniques that use deep convolutional neural networks (DCNNs) have shown capability to drastically reduce the lineaments interpretation life cycle in digital elevation models (DEM), as well as reduce some bias and intrinsic subjectivity. In this work we analyze Ribeira fold belt central region, SE Brazil, and compare manual lineament interpretation with the ones obtained by a DCNN with U-Net architectures trained with data from a nearby area. Furthermore, we examine the biases and limitations regarding the manual method, such as changes in interpretation criteria, motivation and previous experience, the ones related to computers, such as the available data for the training, output noise levels and overfitting. Results show that DCNNs can be used to produce satisfactory lineaments, in fraction of the time required for manual interpretation. However, the search for ideal architectures and hyperparameters may also take a long time. DCNNs are a powerful tool but doesn’t eliminate the interaction with geoscientists that must construct, train, and critically assess its results.