Banca de DEFESA: PRISCILLA AZEVEDO DOS SANTOS

Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.
STUDENT : PRISCILLA AZEVEDO DOS SANTOS
DATE: 13/09/2021
TIME: 09:00
LOCAL: Petrologia
TITLE:

MAPEAMENTO E MODELAGEM DIGITAL DA VARIABILIDADE TRIDIMENSIONAL DE ATRIBUTOS FÍSICO-HÍDRICOS DOS SOLOS DA BACIA DO RIO GUAPI-MACACU-RJ, POR ESTATÍSTICA MULTIVARIADA E ALGORITMOS

KEY WORDS:

Machine Learning. AQP. Multivariate Statistics. Hydropedology. Digital Mapping. Remote Sensing.

PAGES: 155
BIG AREA: Ciências Exatas e da Terra
AREA: Geociências
SUBÁREA: Geologia
SUMMARY:

The knowledge about soil attributes and hydropedology is important for studies that aim
understand the hydrological regime and monitoring water flow, especially in whatersheds,
where the stored and available water content affects both soils environmental functions, as well
as the biodiversity and the sustainability of this natural resource. In Brazil, soil databases have
few information collected about soil water parameters such as basic infiltration speed ratio (bir)
and saturated hydraulic conductivity (Ksat), due to the non-systematic performance of
infiltration tests at soil surveys and the difficulty of measuring such parameters in the
pedosphere`s deepest layers. In this context, it is possible to estimate the bir and Ksat by
associating the granulometric and physicochemical properties of the soil collected in the field,
by means of algorithms for quantitative pedology (Algorithms for Quantitative Pedology -
AQP) and implementation of pedotransfer functions using multivariate regressive analysis and
tree-based machine learning algorithms, capable of modeling them vertically (in profile) and
spatially at the study area. Still, as a way of expanding the information about the studied area
and guaranteeing a more reliable and robust modeling, it is desirable to associate measurable
parameters in the field and laboratory with other relevant information that help the analysis of
hydrographic basins, thus composing the input variables in the models mentioned. This study
suggests the application of variables from numerical modeling of the terrain, obtained through
the Digital Elevation Model (DEM), radiometric data, derived from environmental
aerogeophysics (aeromagnetometry and aerogamaespectrometry) and spectral analysis through
indices related to vegetation, soil and water using images from the Sentinel-2A sensor (spectral
indexes). The quantitative analysis and covariates selection, statistical-descriptive and
multivariate methods were applied, aiming to select potential covariates and the reduction of
dimensionalities and / or multicollinearity in the input variables in the models. Based on the
results obtained, the tree-based models (Random Forest - RF and Regression Trees - RT)
showed better performance in modeling the physical-hydric attributes when compared to the
regressive model to build pedotransfer functions. The multivariate approach using covariates
selection and dimensionality reduction methods allowed the optimized choice of input variables
in the modeling, elimination of multicollinearity problems, obtaining a diversified response to
the evaluated soil layers. The study shows the potential for integrating topographic, pedological
and radiometric data and their contribution to digital soil mapping and modeling, aiming at
understanding the variability of physical-hydric attributes in the studied watershed.

BANKING MEMBERS:
Externo à Instituição - SÍLVIO BARGE BHERING - EMBRAPA
Presidente - 2223668 - HELENA SARAIVA KOENOW PINHEIRO
Externo ao Programa - 1220296 - MARCOS BACIS CEDDIA