Banca de QUALIFICAÇÃO: QUÉSIA SÁ PAVÃO

Uma banca de QUALIFICAÇÃO de DOUTORADO foi cadastrada pelo programa.
STUDENT : QUÉSIA SÁ PAVÃO
DATE: 27/11/2023
TIME: 09:00
LOCAL: online
TITLE:

SOIL TEXTURE PREDICTION OF NATURAL BRAZILIAN AMAZON SOILS BY PROXIMAL SENSORS 

KEY WORDS:

 pXRF, Vis-NIR, random forest 

PAGES: 50
BIG AREA: Ciências Agrárias
AREA: Agronomia
SUBÁREA: Ciência do Solo
SPECIALTY: Fertilidade do Solo e Adubação
SUMMARY:

Proximal sensors are a fast, low-cost, environmentally friendly, and reliable method for detecting element content in soil. While previous research has used these sensors in temperate soils, their use in tropical soils, especially in the Amazon region, is still not well-understood. To address this knowledge gap, this study utilized portable proximal X-ray fluorescence spectrometry (pXRF) and visible and near-infrared spectrometry (Vis-NIR) sensors to predict soil texture in 61 municipalities in the state of Pará, Brazil. The study aimed: i) to investigate the performance of pXRF and Vis-NIR data separately and combined in predicting texture, and ii) to compare the effect of surface horizons, subsurface horizons, and combinations of horizons in predicting soil texture. Soil samples were collected from natural areas with primary or secondary forest cover, with at least 20 years of natural regeneration and approximately 20 ha of cover area, in the 0-20 cm and 80-100 cm layers. Soil texture analysis was carried out using the hydrometer method. A portion of the soil samples were analyzed by pXRF and Vis-NIR sensors in triplicates under laboratory conditions. Soil texture prediction was performed using two machine learning algorithms, the Random Forest (RF) and the Support Vector Machine with Radial Basis Function Kernel (SVM). The RF models showed higher R² values and lower RSME and MAE compared to the SVM models. The R² values with data from pXRF, Vis-NIR, and a combination of sensors were, respectively, for sand (0.89; 0.87 and 0.93), clay (0.92; 0.90 and 0.93), and silt (0.91, 0.67 and 0.93). In general, the models for clay obtained higher R² values compared to sand and silt. The prediction using data from the two sensors obtained lower RMSE and MAE values and higher R² values (sand 93%, clay 93%, and silt 92%) in relation to the best sensor individually (Vis-NIR). The prediction model for sand and clay, using Vis-NIR data, obtained lower errors in both horizons and combinations. The effect of combining horizons was minimally important for the models. The results demonstrate the reliability of using proximal sensors to evaluate soil texture in natural soils in the Amazon, which can help reduce costs and save the time required for soil analyses. 

BANKING MEMBERS:
Presidente - 030.493.086-56 - SILVIO JUNIO RAMOS - ITV
Externo ao Programa - 3600530 - JOAO FERNANDES DA SILVA JUNIOR
Externa à Instituição - EDNA SANTOS DE SOUZA - UNIFESSPA
Externo à Instituição - SÉRGIO HENRIQUE GODINHO SILVA - UFLA