Revista Brasileira de Geomorfologia v. 21, nº 2 (2020) www.ugb.org.br http://dx.doi.org/10.20502/rbg.v21i2.1671 ISSN 2236-5664 MACHINE LEARNING ALGORITHM IN THE PREDICTION OF GEOMORPHIC INDICES FOR APPRAISAL THE INFLUENCE OF LANDSCAPE STRUCTURE ON FLUVIAL SYSTEMS, SOUTHEASTERN - BRAZIL ALGORITMO DE APRENDIZADO DE MÁQUINA NA PREDIÇÃO DE ÍNDICES GEOMÓRFICOS PARA AVALIAÇÃO DA INFLUÊNCIA DA ESTRUTURA DA PAISAGEM EM SISTEMAS FLUVIAIS, SUDESTE – BRASIL Cristiano Marcelo Pereira de Souza Programa de Pós-Graduação em Geografi a, Universidade Estadual de Montes Claros Av. Dr. Ruy Braga, S/N, Montes Claros, Minas Gerais. CEP: 39401-089. Brasil ORCID: 0000-0001-7692-1613 E-mail: [email protected] Natália Aragão de Figueredo Colégio Militar de Juiz de Fora Av. Presidente Juscelino Kubitschek, 5200, Juiz de Fora, Minas Gerais. CEP: 36087-000. Brasil ORCID: 0000-0003-3398-7107 E-mail: [email protected] Liovando Marciano da Costa Programa de Pós-Graduação em Solos Nutrição de Plantas, Universidade Federal de Viçosa Av. Peter Henry Rolfs, s/n, Viçosa, Minas Gerais. CEP: 36570-900. Brasil ORCID: 0000-0001-9581-0783 E-mail: [email protected] Gustavo Vieira Veloso Programa de Pós-Graduação em Solos Nutrição de Plantas, Universidade Federal de Viçosa Av. Peter Henry Rolfs, s/n, Viçosa, Minas Gerais. CEP: 36570-900. Brasil ORCID: 0000-0002-9451-2714 E-mail: [email protected] Maria Ivete Soares de Almeida Programa de Pós-Graduação em Geografi a, Universidade Estadual de Montes Claros Av. Dr. Ruy Braga, S/N, Montes Claros, Minas Gerais. CEP: 39401-089. Brasil ORCID: 0000-0002-3257-7109 E-mail: [email protected] Expedito José Ferreira Programa de Pós-Graduação em Geografi a, Universidade Estadual de Montes Claros Av. Dr. Ruy Braga, S/N, Montes Claros, Minas Gerais. CEP: 39401-089. Brasil ORCID: 0000-0001-8125-6806 E-mail: [email protected] Souza C. M. P. et al. Informações sobre o Artigo Abstract: The Abaeté hydrographic basin was infl uenced by sediments (Proterozoic and Recebido (Received): 19/06/2019 Cretaceous) and by volcanism (Upper Cretaceous). The objective was to identify Aceito (Accepted): whether the drainage confi guration is due to structural factors or by neotectonics. 12/03/2020 We use geomorphic indices (Basin Asymmetry Factor, Stream Length-gradient Index - SL, and Channel Steepness Index - ksn). We elaborate longitudinal profi les Keywords: in the rivers of sub-basins. We apply the Random Forest (RF) Machine Learning Hydrographic Basin; Structural algorithm in the prediction of the SL and k indices, with the selection of relevant Drainage Control; Random Forest; sn covariates. The RF was effi cient in predicting, with better performance in the k Morphotectonic Analysis; sn (R2 0.38), and indicating the areas of infl uence of the indices. The highest values Palavras-chave: of the indices are in zones of lithological contact with diff erent resistances, Bacia Hidrográfica; Controle favoring a sharp change in channel slope (knickpoints). In these zones, there Estrutural da Drenagem; Random is also a predominance of sub-basins tilted, and the longitudinal profi les of the Forest; Análise Morfotectônica; rivers show uplift or subsidence. Therefore, structural factors conditioned the drainage of the basin. Resumo: A bacia hidrográfi ca de Abaeté foi infl uenciada por sedimentação (Proterozóico e Cretáceo) e processo vulcânicos (cretáceo superior). O objetivo foi identifi car se há um controle estrutural dos sistemas fl uviais relacionado ao vulcanismo ou por neotectônica. Utilizamos os índices geomórfi cos (Fator de Assimetria de Bacia, Stream Length- gradient Index – SL, e Channel Steepness Index - ksn); elaboramos perfi s longitudinais nos rios de sub-bacias. Aplicamos o algoritmo de Aprendizado de Máquina Random Forest (RF) na predição dos índices SL e ksn, com 2 seleção de covariáveis importantes. O RF foi efi ciente na predição, com melhor performance em ksn (R 0,38), e indicando as áreas de infl uência dos índices. Os valores mais altos dos índices estão em zonas de contato litológico com diferentes resistências, favorecendo mudanças acentuadas de declividade no canal (knickpoints). Nessas zonas, também há predominância de sub-bacias basculhadas, e os perfi s longitudinal dos rios mostram elevação ou subsidência. Portanto, fatores estruturais condicionam os sistemas fl uviais da bacia. 1. Introduction Normalized channel steepness index (ksn) (WOBUS et al., 2006), has the advantage of inserting variables The river systems are sensitive to the evolution according to the drainage characteristics (CASTILLO of the relief, generating changes in the morphology et al., 2014). There is also the Basin Asymmetry Factor of the rivers (FONT et al., 2010), abrupt changes in index, capable of identifying tilted hydrographic basin slope in the longitudinal profi le of the river (AMBILI (HARE & GARDNER, 1985). Studies demonstrate & NARAYANA, 2014), and asymmetric basins (EL the effi ciency of these methods in diff erent geological HAMDOUNI et al., 2008). Therefore, studies use contexts (GARROTE et al., 2008; FONT et al., 2010; analysis of the drainage network to identify structural MONTEIRO et al., 2010; CYR et al., 2014; SOUZA & control (KALE et al., 2014), and also deformations by PEREZ FILHO, 2017), providing a quick assessment neotectonics (GARROTE et al., 2008; MONTEIRO et for large areas, and with quantitative data. al., 2010; ALVES et al., 2019). The geographic information system (GIS) com- Visual assessment of drainage characteristics is bined with topographic data from digital elevation essential, but quantitative methods are complementary models (DEM), represent an advance in the extraction and make the analysis more objective, and geomorphic of geomorphic indices (TROIANI et al., 2014; ALVES indices perform this function. Among these indices, et al., 2019). However, some data may have limitations the Stream Length-gradient Index method (SL), iden- in cartographic representation, for example, the SL and tifi es abrupt changes in slope along the profi le (knick- k indices are represented by points (MONTEIRO et points) (HACK, 1973). Based on the SL indices, the sn al., 2010) or by drainage network (E HAMDOUNI Rev. Bras. Geomorfol. (Online), São Paulo, v.21, n.2, (Abr-Jun) p.365-380, 2020 Machine Learning Algorithm in the Prediction of Geomorphic Indices for Appraisal the Infl uence of Landscape et al., 2008; CASTILLO et al., 2014). Therefore, to of reactivation in the Quaternary, being able to change advance the cartographic representation, studies use the morphology of the current drainage. Some studies kriging methods, creating homogeneous zones of in- demonstrate reactivation processes in the interior of fl uence of the indices (knickzones) (P‐P et al., Brazil, for example, in the region of Poços de Caldas- 2009; FONT et al., 2010). However, kriging effi ciency Minas Gerais, by reactivation of the Triassic shear zones depends on point density, spatial dependence and aniso- (GROHMANN et al., 2007; DORANTI-TIRITAN et tropy (OLIVER & WEBSTER, 1990); therefore, these al., 2014). characteristics that can be limiting, justify testing new This study aims to identify evidence of structural methods of prediction. and/or neotectonic control using geomorphic indices Recently, Machine Learning algorithms (ML) and prediction methods of indices by the Machine has gained prominence in spatial prediction studies. Learning. The study area is the Abaeté hydrographic In general, ML has the advantage of performing spa- basin, located in the central portion of the state of Minas tialization from point values (variable), based on a Gerais, where there was an intense process of sedimen- database in raster format (predictive covariates), and tary deposition during the Proterozoic and volcanism using some statistical algorithm (KUHN & JOHNSON, in the Cretaceous. 2013). Studies show the effi ciency of ML in the area of geology (CRACKNELL et al., 2014), geomorphology 1.1 Physiography and geological framework (MOHAMMADY et al., 2019) and pedology (SOUZA et al., 2018; GOMES et al., 2019). Besides, ML can The study area is the Abaeté hydrographic basin, perform better than the kriging method (CLODOAL- located in southeastern Brazil (Minas Gerais state), VES, 2019), especially when the variables do not have between the coordinates -17 ° 85 ’to -19 ° 36’ latitude S spatial dependence; however, there are no studies on the -45 ° 21 ’to -46 ° 31’ longitude W, with an area of 5,824 prediction of geomorphic indices with ML. km2 (Figure 1). The climatic domain is Cwa (Köppen Most studies with geomorphic indices are on classifi cation), with an average annual precipitation of tectonically active margins, focusing on neotectonics 1,700 mm. (SEEBER & GORNITZ, 1983; TROIANI et al., 2014). The basin has Proterozoic and Cretaceous sedi- However, in Brazil, passive margin, some studies attest mentary rocks and the infl uence of magmatic activity to the occurrence of neotectonic processes, by abrupt during the Cretaceous (Figure 1). In general, in Pro- changes in river direction, knickpoints, and straight terozoic lithologies, occur depressions and dissected rivers (ETCHEBEHERE et al., 2004; SOUZA et al., areas. On the other hand, the Cretaceous rocks form 2011; DORANTI-TIRITAN et al., 2014), with some plateaus, forming two morpho sculptural compart- characteristics quantifi ed by geomorphic indices. ments, one preserved surface (Mata da Corda Group), In Brazil, major tectonic events occurred mainly in the other dissected (Areado Group) (OLIVEIRA & the separation of the South