Banca de QUALIFICAÇÃO: FABRICIO DO CARMO FARIAS

Uma banca de QUALIFICAÇÃO de MESTRADO foi cadastrada pelo programa.
STUDENT : FABRICIO DO CARMO FARIAS
DATE: 30/06/2022
TIME: 14:00
LOCAL: Videoconferência
TITLE:

Soil mapping and use of machine learning algorithms in Tracuateua (PA)

KEY WORDS:

DSM, accuracy, Radom Forest,  Recursive Partitioning.

PAGES: 45
BIG AREA: Ciências Agrárias
AREA: Agronomia
SUBÁREA: Ciência do Solo
SUMMARY:

Difficulties in access in areas of the Amazonian biome have limited economic and agro -environmental planning analysis in the northern region of Brazil. In this sense, the soil digital mapping tools (MDS), notably, machine learning techniques have helped to obtain quality soil maps and known accuracy. And considering the variability of models available today, it is important to evaluate their performance in relation to the data set and environmental variables involved in the digital soil mapping. This work aimed to perform the digital mapping of soils of the municipality of Tracuateua/PA, on a scale 1: 100,000, evaluate different methods, and determine the set of significant environmental covariables that represent the soil formation factors and explain the local pedological variance. The study evaluated the performance of two machine learning algorithms (Radon Forest - RF and Recursive Partitioning - RPART) for soil mapping in the municipality of Tracuateua/PA, Northeast Paraense, Eastern Amazon. 44 geomorphometric covariables were used from a digital elevation model - MDE, obtained from Palsar, local geology and Landsat 8 satellite vegetation rates, using GIS 2.3.2 saga software and qgis 3.16.11. In the SFWARE R 4.2.0, covariable selection was selected by the collinearity elimination algorithm and successively. 2 machine learning algorithms were tested as classifiers, and the analyzes were conducted using the Dplyr, Tidyr, Forcats, Tibble, here, Caret, Earth, Grepel, GGPLot2, SF, Readr, Janitor, Future and Future packages. Preliminary results indicate that the two algorithms performed similar performance (Kappa 0.44-0.66). Soil classes: (Yellow Latosol - LA, Yellow Argisol - PA, quartzenic neosols - RQ) and Hydromorphic Plains (Hamable Gleissols - GX) obtained greater agreement with the conventional map, with greater performance of the algorithms. In general the global agreement rate obtained by the map algebra showed that the result was satisfactory, with 49-60 % agreement between the conventional soil map and maps produced by machine learning. The most disagreement areas in MDS occurred in the PA unit due to low correlation with environmental variables. The performance of the models was satisfactory, with good agreement with the conventional soil map, this highlights the importance of MDS as a potential complementary tool to assist soil mapping in difficult areas in Brazil, especially in regions, as in the Amazon

BANKING MEMBERS:
Presidente - 3600530 - JOAO FERNANDES DA SILVA JUNIOR
Externo ao Programa - 1830598 - DANIEL PEREIRA PINHEIRO
Externa à Instituição - IOLANDA MARIA SOARES REIS - UFOPA
Externo à Instituição - DANIEL DE BORTOLI TEXEIRA