Predictive lithological mapping through machine learning methods: a case study in the Cinzento Lineament, Carajás Province, Brazil

Journal of the Geological Survey of Brazil

Endereço:
SBN Quadra 2 Bloco H - 1° andar - Asa Norte
Brasília / DF
70040-904
Site: https://jgsb.sgb.gov.br/index.php/journal/index
Telefone: (61) 2108-8400
ISSN: 2595-1939
Editor Chefe: Evandro Luiz Klein
Início Publicação: 15/05/2018
Periodicidade: Quadrimestral
Área de Estudo: Multidisciplinar, Área de Estudo: Multidisciplinar

Predictive lithological mapping through machine learning methods: a case study in the Cinzento Lineament, Carajás Province, Brazil

Ano: 2019 | Volume: 2 | Número: 1
Autores: I. S. L. Costa , F. M. Tavares , J. K. M. . Oliveira.
Autor Correspondente: I. S. L. Costa | [email protected]

Palavras-chave: random forest, cinzento lineament, machine learn, airborne geophysics.

Resumos Cadastrados

Resumo Inglês:

The Cinzento Lineament (Carajás Mineral Province) represents a complex deformational system with great associated mineral potential, mainly for IOCG deposits. However, the tropical vegetation of the Amazon rainforest considerably limits the number of outcrops available for systematic geological mapping. Therefore, the use of remote data such as airborne geophysics and remote sensing is essential to provide a reliable geological map. The airborne magnetometric data to define lithological units and its boundaries is a challenge, especially in regions with low magnetic latitude and/or remanent magnetization. In this work, we proposed an approach using Magnetization Vector Inversion (MVI) to map the distribution of the magnetic susceptibility, in order to replace techniques such as pole reduction and total gradient. We applied the Random Forest algorithm (supervised Machine Learning algorithm) to recognize patterns in remote data and improve the current mapped lithological units. With 1400 training samples (2.5% of the total samples), we produced two Predictive lithological maps: a first with remote data only and a second with remote data and spatial coordinates. We evaluate the advantages and disadvantages of each Predictive map, and we conclude that both maps need to be analyzed together for the refinement of the current geological map. These predictive maps represent a powerful tool to combine remote data to improve current geological maps, or even generate the first-pass geological map for regions with scarce geological knowledge.