Desenvolvimento e validação de modelos de redes neurais para estimativa de evapotranspiração de referência no Paraná

• Autor(a) principal: Postal, Adriana
• Data de Publicação: 2024
• Tipo de documento: Tese
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações do UNIOESTE
• Texto Completo: https://tede.unioeste.br/handle/tede/7240
• topic: Redes neurais na agricultura; Manejo da irrigação; Dados meteorológicos; Neural networks in agriculture; Irrigation management; Meteorological data; RECURSOS HIDRICOS E SANEAMENTO AMBIENTAL

This thesis addressed the estimation of reference evapotranspiration (ETo) in the agricultural context, focusing on efficient irrigation management. The main objective was to develop and validate models based on Multilayer Perceptron Artificial Neural Networks (MLP ANNs) for ETo prediction in the state of Paraná, using data from the National Institute of Meteorology (INMET) and a producer in the western region. The research began with a review of irrigation management and the Penman-Monteith method for ETo estimation, as well as the use of ANNs, highlighting challenges and relevant approaches. The methodology involved the collection and organization of meteorological data from INMET, essential for model training, and the development and training of ANNs using two optimizers (SGD and Adam), exploring meteorological variables as input to predict ETo, along with model validation using producer data. The analysis considered statistical metrics, including Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Coefficient of Determination (R2). The results revealed that the configurations RNA9 (with 8 input variables) and RNA10 (with 4 input variables) showed promising performance in ETo estimation. Validation with producer data highlighted the practical relevance of the models. The MAE, MSE, RMSE, and R2 metrics indicated efficacy, while the Kolmogorov-Smirnov test contributed to evaluating the predictions’ adherence to the statistical behavior of real data. The metric results showed that the best configuration was RNA9, with a mean absolute error of only 0.01 mm/day with both optimizers, and the configuration with fewer variables was RNA10, with a mean absolute error of only 0.03 mm/day, also with both optimizers. With R2 ranging between 0.99 and 1, it is possible to affirm that the models suited the data from the state of Paraná, even when confronted with producer data not used in training. Suggestions for future research include developing an integrated system for automated collection of meteorological data from INMET and implementing an irrigation management application based on the most efficient models. At the end of this work, it was possible to develop and validate ANN models for simplified and accurate ETo prediction in agriculture, promoting an accessible approach to optimize irrigation management in the state of Paraná.