Banca de DEFESA: DOUGLATH ALVES CORRÊA FERNANDES

Uma banca de DEFESA de DOUTORADO foi cadastrada pelo programa.
STUDENT : DOUGLATH ALVES CORRÊA FERNANDES
DATE: 31/10/2024
TIME: 13:30
LOCAL: videoconferência
TITLE:

Digital Mapping of Soil Chemical Attributes in an Oil and Gas Production Area in the Northern Region of Recôncavo Baiano: A Case Study in the Semi-Arid Region of Sátiro Dias, Bahia.

KEY WORDS:

Covariate selection, machine learning, digital soil mapping.

PAGES: 90
BIG AREA: Ciências Agrárias
AREA: Agronomia
SUBÁREA: Ciência do Solo
SUMMARY:

Exploring new technologies and approaches in soil mapping that allow the production of high-resolution maps at low cost is a continuous challenge. In this context, using approaches that focus field observations on smaller areas can yield better results compared to methods that use the entire area for sampling. This study presents and tests the approach known as the Reference Area, which allows the generation of maps of soil chemical attributes with lower costs, high spatial resolution, and high accuracy over large territorial extensions. This study is divided into two chapters. The first chapter aims to map the spatial variability of SOCS at depths of 0-30 cm and 0-100 cm in the semi-arid region of Bahia, while the second chapter focuses on modeling and mapping the percentage of sodium saturation (PST) at five depths using machine learning algorithms and testing three covariate selection methods for a study area located in the municipality of Sátiro Dias in the northern region of Bahia. In both chapters, two different approaches were tested and compared: the Reference Area (RA) and Total Area (TA) approaches. Three machine learning models were used: Random Forest (RF), Cubist, and Support Vector Machine (SVM), along with three covariate selection methods: all covariates (TC), Recursive Feature Elimination (RFE), and Expert Knowledge (EK). The study area is located in the northern region of Bahia, Brazil, and covers a total area of 900 km², with a smaller area of 200 km² pre-selected as the reference area. The database consists of 124 sampling points divided into 2 depths for SOCS and 5 depths for PST. In the first chapter, it became evident that the choice of model and covariate selection technique considerably affects carbon stock prediction, with EK standing out as the best covariate selection method. Another important result was regarding the use of coordinates as a covariate: using latitude and longitude improves model accuracy but introduces errors in the visual assessment of spatial predictions, resulting in clear artifacts in the maps. Furthermore, the results show that, for the study area, the choice of reference area did not produce results equivalent to the total area. In the second chapter, the results highlighted the efficiency of the machine learning algorithms Cubist and Random Forest in mapping PST at different soil depths in the study area. Although the approach based on the Reference Area proved promising for modeling and extrapolation, its results did not surpass those obtained with the total area. Covariate selection was also successful, with the Recursive Feature Elimination (RFE) technique standing out for its effectiveness.

COMMITTEE MEMBERS:
Externo ao Programa - 2365983 - ANDRE LUIS OLIVEIRA VILLELA - nullExterno à Instituição - GUSTAVO SOUZA VALLADARES - UFPI
Presidente - 1220296 - MARCOS BACIS CEDDIA
Interno - 1060711 - MARCOS GERVASIO PEREIRA
Externa à Instituição - SANDRA SANTANA DE LIMA - UFRRJ
Externo à Instituição - WALDIR DE CARVALHO JUNIOR - EMBRAPA