Banca de QUALIFICAÇÃO: ROMARIO DE SOUSA ALMEIDA

Uma banca de QUALIFICAÇÃO de DOUTORADO foi cadastrada pelo programa.
DISCENTE: ROMARIO DE SOUSA ALMEIDA
DATA: 14/04/2025
HORA: 08:00
LOCAL: Online, no Google Meet
TÍTULO:

FROM SEEDLING DETECTION TO CLIMATE FORECASTING: ARTIFICIAL INTELLIGENCE APPLIED TO SMART GREENHOUSE MONITORING

PALAVRAS-CHAVES:

computer vision; deep learning; YOLO; machine learning; fuzzy logic; regression model; mathematical modeling; automation; agricultural monitoring; smart agriculture.

PÁGINAS: 82
GRANDE ÁREA: Ciências Agrárias
ÁREA: Engenharia Agrícola
SUBÁREA: Construções Rurais e Ambiência
RESUMO:

The present study explores the application of artificial intelligence techniques for intelligent monitoring in greenhouses, focusing on two main challenges: the automatic detection of lettuce seedlings and the forecasting of the internal microclimate of the greenhouse. The first article aims to fine-tune and evaluate the performance of YOLOv5 to YOLO11 models for detecting lettuce seedlings in a greenhouse environment. Image datasets were organized for training, validation, and testing, with bounding box annotations for each plant and empty cell (non-germinated), assessing the configurations from YOLOv5 to YOLO11. Over 15 days, 230 images were collected, totaling 3,450 images, divided in an 8:1:1 ratio. YOLOv8 was the fastest model, with an inference time of 2.9 milliseconds (ms), followed by YOLOv6 at 3.3 ms. YOLOv5 achieved the highest seedling detection performance compared to the other evaluated models, with a precision of 0.9992, a recall of 0.9987, an mAP@0.5 of 0.9950, an mAP@0.5:0.95 of 0.9665, and an F1-Score of 0.9989. Therefore, YOLOv5 can be recommended for the precise detection of lettuce seedlings in greenhouse environments. In the second article, data normalization of temperature and humidity is proposed to develop a fuzzy climate forecasting model for the greenhouse as an alternative to the statistical regression model. Two datasets will be used: the first consists of temperature and relative humidity data collected inside the greenhouse, and the second comprises historical data obtained from the Climatological Station. Missing values and outliers will be treated. Then, four normalization methods will be evaluated: Min-Max, Z-Score, Absolute Maximum, and Sigmoidal Normalization. The best method will be selected based on a comparative analysis of each technique’s impact on predictive model performance, following the criteria: data distribution preservation, improvement in model performance, and interpretability of results. To develop the multiple linear regression models, the temperature and relative humidity data from the meteorological station were considered independent variables, while the dependent variables are the temperature and relative humidity data collected inside the greenhouse. Initially, an analysis of variance will be performed using the F-Test (p < 0.05) to assess the statistical significance of the coefficients. Subsequently, regression models will be adjusted to forecast the microclimate. The fuzzy inference systems will be developed by defining the membership functions for the input variables (data from the Meteorological Station) and output variables (data collected inside the greenhouse). The database will be randomly split into two subsets, with 80% used for system development and 20% for validation. Two inference methods will be applied: Mamdani and Sugeno. All defuzzification methods available in MATLAB will be tested. The models’ performance will be compared using accuracy and error metrics. Thus, this research contributes to the modernization of agriculture by promoting intelligent solutions for monitoring and automating the management of climate conditions and seedling production within the greenhouse.

MEMBROS DA BANCA:
Externo ao Programa - VALTER CARVALHO DE ANDRADE JUNIOR - DAG/ESAL (Membro)
Interno - TADAYUKI YANAGI JUNIOR (Membro)
Externo à Instituição - Jayane Karine Pereira de Araújo - UFPB (Suplente)
Externo ao Programa - FELIPE SCHWERZ - DEA/EENG (Membro)
Presidente - ALESSANDRO TORRES CAMPOS (Membro)