Banca de QUALIFICAÇÃO: CAIO VICTOR MACEDO PEREIRA

Uma banca de QUALIFICAÇÃO de MESTRADO foi cadastrada pelo programa.
STUDENT : CAIO VICTOR MACEDO PEREIRA
DATE: 04/12/2025
TIME: 08:30
LOCAL: Natal
TITLE:

ARTIFICIAL INTELLIGENCE IN LAND USE AND COVER CLASSIFICATION IN SEMI-ARID REGIONS

KEY WORDS:

Remote sensing, GIS, Caatinga, Supervised Algorithms

PAGES: 40
BIG AREA: Engenharias
AREA: Engenharia Sanitária
SUMMARY:

Improving the accuracy of land use and land cover classification in semi-arid regions is crucial due to the environmental vulnerability of these regions and the lack of studies integrating satellite images of different spectral resolutions. In the Caatinga biome, where temporal variability associated with climate results in unique spectral responses, advanced machine learning techniques are particularly useful for improving the accuracy of land use and land cover mapping, in conjunction with other aspects of the classification process. The study evaluated five classifiers: Random Forest (RF), K-Nearest Neighbors (k-NN), Quadratic Discriminant Analysis (QDA), and two artificial neural networks, a Multilayer Perceptron (MLP, implemented with scikit-learn) and a Dense Neural Network (DNN, implemented with Keras), applied to multispectral images (PlanetScope PSB.SD and Sentinel-2A MSI) in the periods of April (rainy) and October (dry). A classification by photo interpretation performed from PlanetScope satellite images served as a reference map with seven thematic classes. The classifiers were implemented in Python and trained/tested on a pixel-by-pixel basis for each sensor/season combination. The highest accuracies were obtained by neural network architectures: RND on PlanetScope (wet period) achieved Global Accuracy = 91.12% and MLP on PlanetScope (wet period) achieved 91.01%. Random Forest and k-NN showed intermediate performance, and ADQ obtained the worst results. Cross-temporal application (training in wet conditions, application in dry conditions) showed low transferability (≈45% in the k-NN test). Although PlanetScope presented greater geometric detail, the accuracy of the best classifiers was similar between sensors. In view of the confusions observed (exposed soil, pasture, crops in the dry season) and the limitations of pixel-wise approaches, the use of multitemporal series, multisensor fusion, object-oriented/segmentation techniques, and convolutional architectures is recommended to incorporate spatial and phenological context, aiming to increase the accuracy and robustness of mapping in semi-arid environments.

COMMITTEE MEMBERS:
Presidente - 3060504 - CARLOS WILMER COSTA
Interno - 2411669 - JONATHAN MOTA DA SILVA
Interno - ***.898.694-** - PAULO VICTOR DO NASCIMENTO ARAÚJO - UFRN