Banca de QUALIFICAÇÃO: EDUARDA CAVALCANTE AMORIM

Uma banca de QUALIFICAÇÃO de MESTRADO foi cadastrada pelo programa.
STUDENT : EDUARDA CAVALCANTE AMORIM
DATE: 27/02/2025
TIME: 09:00
LOCAL: On line
TITLE:

DIGITAL  SOIL ORGANIC CARBON MAPPING BASED ON DEEP LEARNING FROM REMOTE SENSING DATA IN REGIONAL-SCALE

KEY WORDS:

Machine Learning; Climate change; Pedometry, Amazon

PAGES: 40
BIG AREA: Ciências Agrárias
AREA: Agronomia
SUBÁREA: Ciência do Solo
SUMMARY:

The increase in carbon dioxide (CO₂) emissions into the atmosphere has become notably worrying in recent decades, making it crucial to monitor soil organic carbon stocks in the Amazon region, which represents a large carbon reservoir worldwide. Accurate mapping of organic carbon stocks at a regional scale presents a series of complex challenges due to spatial heterogeneity, variability in soil formation and degradation processes, and the influence of factors such as vegetation, climate, and soil management. Accurate estimation over large geographic areas requires the integration of multiple data sources and the application of advanced analytical methods. The research aims to evaluate the efficiency of deep learning algorithms for determining soil organic carbon stocks (SOC) in the municipality of Capanema, in northeastern Pará. The study was carried out between July 2022 and July 2024, in an area belonging to the municipality of Capanema-PA, soil samples were obtained in a sampling grid of 1 point every 1,800 m, in the 0-20 cm layer. The images were obtained from the Sentinel 2 satellite, Copernicus Digital Elevation Model, Google Earth Engine and IBGE data source, following the dates of soil collections, then they were processed in the QGIS software and R programming language and then the most important environmental covariates were selected for determining the ECOS, through three selection algorithms: Boruta, recursive feature elimination (RFE), SHapley Additive exPlanations (SHAP). The data were divided and intended for training and validation of the chosen models, 70% and 30% respectively, being the Randon Forest (RF), Artificial Neural Network (ANN), Support Vector Machine (SVM) and Deep Neural Network (DNN). The results obtained through the models must be submitted to the Root Mean Square Error (RMSE), Coefficient of Determination (R²) and Lin's Concordance Correlation Coefficient (LCCC) indices to verify their performances and define the best model. As partial results, most of the sampled points were in pasture areas with a coefficient of variation of 49%, considered average. Environmental covariates submitted 74 environmental covariates in Spearman's correlation analysis and 28 covariates that presented low correlation with ECOS were eliminated. After this data mining step, 46 covariates were then submitted to the RFE selection algorithm, which assembled a scenario of the most important covariates for determining soil organic carbon stock, with 44 covariates that were submitted to machine learning with the RF algorithm, of which the most important for determining ECOS are related to relief and vegetation indices, such as TX, RSP, EPRI and MNDWI. These results are important for choosing the most appropriate covariates and machine learning models for digital mapping of organic carbon stock in the Amazon.

COMMITTEE MEMBERS:
Presidente - 3600530 - JOAO FERNANDES DA SILVA JUNIOR
Interna - 2121042 - SUZANA ROMEIRO ARAUJO
Externo à Instituição - GENER TADEU PEREIRA - UNESP