The role of airborne geophysical data as covariate to map soil classes: study case - southern portion of the Paraná sedimentary basin
Visualizar/ Abrir
Data
2023-09-19Primeiro coorientador
Schenato, Ricardo Bergamo
Primeiro membro da banca
Pinheiro, Helena Saraiva Koenow
Segundo membro da banca
Horst, Taciara Zborowski
Metadata
Mostrar registro completoResumo
The southern Parana Sedimentary Basin has a unique context of soil parental material contrasts and a complex
landscape. Airborne geophysical data, such as gamma-ray spectrometry and magnetometry, have great potential
to represent soil parent material, helping spatial predict tropical soil classes. Because of that, this study used
morphometric and geophysical data in different combinations to test and compare different machine learning (ML)
algorithms to investigate the impact of airborne geophysical data on soil class prediction. Twelve unique models
were evaluated, encompassing random forest (RF), support vector machine (SVM), and artificial neural networks
(ANN) ML algorithms through four model formulations of covariate predictors - morphometric (MORPHO),
gamma-ray spectrometry (GAMMA), magnetometry (MAG), and a completed set containing all covariates (ALL).
The MOPHO served as a baseline for subsequent comparative analyses with the individual geophysical dataset
and their addition. Additionally, feature selection was employed in three steps (variance near zero, removal by
correction, and removal by importance), the importance of the covariates in predictive models was analyzed, and
the entire modeling framework was submitted to 100 interactions. The study was conducted with 1759 samples
categorized into 13 soil classes (2nd taxonomic level) of the Sistema Brasileiro de Classificação de Solo, covering
an area of approximately 2700kmš. The results showed that the morphometric covariates, particularly elevation,
were consistently crucial predictors, highlighting the importance of relief. The anomalous potassium (Kd) and
eTh/K covariates, related to weathered soils, were the most important gamma-ray spectrometry covariates. The
covariates x horizontal derivative (gx) and anomalous magnetic field (cma) were the most important magnetometry
covariates, ruled structures of the relief. The RF algorithm predicted soil classes with the highest means of Overall
Accuracy values. The SVM and ANN algorithms demonstrated a moderate performance compared; however, none
of the maps predicted by the SVM algorithm included the entire set of target variables present in the model fitting.
The ALL model showed the highest performance regardless of the algorithm employed. The MORPHO model was
often the second best. In this sense, the hypothesis that airborne geophysical data effectively can help soil parental
material representation in the Digital Soil Mapping (DSM) approaches was corroborated, thus enhancing the usual
morphometric covariates employed. Results highlight the importance of time (age) in local soil formation, which
is often challenging to represent, emphasizing the common diversity of the soil types and Mesozoic geologicalstratigraphic contrast. This research highlights the structural influence of relief on low-range broad hills, showing
a directional alignment in the NW-SE direction associated with the varied Acrisols. It also distinguishes between
PV ("red" Acrisols) in the northern region, related to igneous rocks, and PV in the central-southern region, connected to sedimentary rocks. Moreover, it identifies the PBAC ("blackish gray" Acrisols) as explicitly associated
with a distinct geophysical domain within the Sanga do Cabral Formation. The study underscores the potential
of integrating geophysical data in DSM, emphasizing the importance of a comprehensive approach that combines
diverse data sources and algorithms, and the world of research questions can emerge from the initial use of legacy
data.
Coleções
Os arquivos de licença a seguir estão associados a este item: