Banca de DEFESA: Ismael Cavalcante Maciel Junior

Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.
DISCENTE : Ismael Cavalcante Maciel Junior
DATA : 28/02/2023
HORA: 08:00
LOCAL: Google Meet (videoconferência)
TÍTULO:

Detection of maize areas through geo-object oriented analysis using orbital multi-sensors on the platform Google Earth Engine

PALAVRAS-CHAVES:

GLCM Textures, Landsat-8, MODIS, PlanetScope, Random Forest, Sentinel 2, SNIC segmentation.

PÁGINAS: 96
GRANDE ÁREA: Ciências Exatas e da Terra
ÁREA: Geociências
SUBÁREA: Geofísica
ESPECIALIDADE: Sensoriamento Remoto
RESUMO:

The state of Mato Grosso is the largest producer of corn in the country, with the predominance of cultivation concentrated in the second harvest. Given the desire for data with greater accuracy and agility, agricultural intelligence needs to adapt its research and search for new technologies such as the use of satellites in remote sensing and geographic information systems. Therefore, this work aimed to map the second-crop corn in the municipality of Canarana-MT through geo-object oriented analysis with different spatial, spectral, temporal and radiometric resolutions, using the sensors: Sentinel-2A/B (MSI), Landsat-8 (OLI), Terra/Aqua (MODIS) and PlanetScope (PS). To achieve these objectives, the IBGE cartographic base of the Canarana municipal boundary was obtained, and later uploaded to the Google Earth Engine Assets (GEE). Filters were applied to the acquired images, selecting images up to 0.35% with cloud occurrence and the median of the images in the period from April to May 2022. Besides the original bands, four vegetation indices were added (NDVI, EVI, PCI and PCEI), and in order to achieve more efficient results in the segmentation step, textural features extracted by means of Gray Level Co-occurrence (GLCM) were employed. After the mentioned procedure, a principal component analysis (PCA) was performed in order to reduce the dimensionality of the data. The images went through the segmentation stage, qualified by applying the Simple Non-Iterative Clustering (SNIC) algorithm. In the classification step, the Random Forest (RF) that is considered a machine learning algorithm was applied. After processing and classifying the images and making the thematic maps of the corn areas, we proceeded to the analysis of the numerical confusion matrix implemented in GEE and validation statistics using 2,200 field samples. From these procedures, the classified data generated the confusion matrices with Global Accuracy (GA) ranging from 86.41% to 88.65%. The OLI image presented the best results, with a GA of 88.65% and 84.61% for the Kappa index, with Producer Accuracy (PA) for the analyzed classes between 65.16% and 91.53%, being the lowest for the Other Land Uses class and the highest corresponding to the second-crop corn class. The Consumer Accuracy (CA) of the aforementioned class was 91.98%. Thus, it is concluded that the GEOBIA methodology using the combination SNIC + GLCM with the RF classifier on the GEE platform was satisfactory.

MEMBROS DA BANCA:
Presidente - 265126001 - CARLOS ANTONIO DA SILVA JUNIOR
Interno - 131916001 - RIVANILDO DALLACORT
Externo à Instituição - CÁCIO LUIZ BOECHAT - UFPI