Reamostragem em redes neurais: uma abordagem alternativa aos métodos tradicionais de interpolação espacial para modelagem de superfícies em áreas agrícolas

• Ano de defesa: 2024
• Autor(a) principal: Silva, Wilker de Jesus
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Uberlândia; Brasil; Programa de Pós-graduação em Agricultura e Informações Geoespaciais
• Palavras-chave em Português: rede neural artificial; artificial neural network; divisão de dados; data splitting; reamostragem; resampling; leave-one-out; predição de intervalo; interval prediction; CNPQ::CIENCIAS EXATAS E DA TERRA::GEOCIENCIAS; CNPQ::CIENCIAS AGRARIAS::AGRONOMIA; Agronomia; Cultivos agrícolas; Estações agrícolas experimentais; Inovações agrícolas; Redes neurais (Computação); ODS::ODS 2. Fome zero e agricultura sustentável - Acabar com a fome, alcançar a segurança alimentar e melhoria da nutrição e promover a agricultura sustentável.
• Link de acesso: https://repositorio.ufu.br/handle/123456789/41633; http://doi.org/10.14393/ufu.di.2024.118

Artificial Neural Networks (ANNs) have been employed in a wide range of applications, representing a highly potential mechanism for data analysis and problem-solving across various fields of knowledge. In surface modeling, ANNs play the role of a spatial interpolation method. However, when employing neural networks, predictions of altitude values do not provide their corresponding uncertainties. In this contribution, this research provides an enhancement of ANN predictions through an innovative approach by applying a resampling method to predict altitude intervals instead of a single estimate, as commonly done by conventional techniques. A Multilayer Perceptron (MLP) network was used to predict altitude intervals based on the coordinates of field-collected points using Real Time Kinematic (RTK) positioning. The network was trained and validated using the Repeated Leave-One-Out Cross-Validation (RLOOCV) resampling method, an extension of the classic Leave-One-Out Cross-Validation (LOOCV) method, which, through multiple iterations, captures the randomness associated with the neural network, including factors like architecture, initialization, and learning procedure. The performance metrics demonstrated satisfactory results in altitude estimation, presenting consistent Root Mean Square Error (RMSE) values, with the global average RMSE, maximum RMSE and minimum RMSE of 0.081 m (± 0.002), 0.520 m e 0.027 m, respectively. Although this methodology demonstrated satisfactory performance, spatial analysis revealed challenges in generalizing to slope, valley, and areas with abrupt variations in terrain inclination.