Análise preditiva de dados geoespaciais para mitigação de desastres climáticos em propriedades rurais
Carregando...
Arquivos
Data
Autores
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual de Ponta Grossa
Resumo
A agricultura, no estado do Paraná, destaca-se pela alta adoção tecnológica e produtividade, porém apresenta vulnerabilidade estrutural crônica frente à variabilidade e aos extremos climáticos. Nesse contexto, esta dissertação teve como objetivo geral desenvolver e validar um Sistema de Alerta Precoce (SAP) baseado em algoritmos de Machine Learning para a predição espacial e antecipação de quebras de safra nas principais culturas de sequeiro da região (soja, milho, trigo e feijão). A metodologia fundamentou-se na integração espaço-temporal de séries meteorológicas diárias do Instituto Nacional de Meteorologia (INMET) com o histórico anual de rendimento agrícola do Instituto Brasileiro de Geografia e Estatística (IBGE), abrangendo o período de 2011 a 2024. Para mitigar o ruído estatístico de análises anuais contínuas, propôsse uma engenharia de atributos estritamente acoplada aos calendários fenológicos locais. A validação algorítmica descartou métodos aleatórios, adotando a divisão cronológica estrita (Out-of-Time Validation), utilizando as safras de 2011 a 2020 para o treinamento e o período climaticamente anômalo de 2021 a 2024 para o teste cego. O desbalanceamento de classes, inerente aos desastres naturais, foi mitigado por meio da técnica de sobre amostragem sintética SMOTE, conjugada à Calibração de Limiar (Threshold Optimization). Os resultados evidenciaram a ocorrência do "Paradoxo da Acurácia", no qual arquiteturas complexas (como as Redes Neurais MLP) falharam na detecção de anomalias quando avaliadas pelo erro global. Ao reposicionar a métrica de Sensibilidade (Recall) como critério de sucesso operacional, o Modelo Support Vector Machine (SVM) com kernel RBF destacou-se na identificação precoce de quebras na cultura da soja, enquanto o Algoritmo Random Forest apresentou a melhor relação de trade-off para o trigo. A análise de importância de variáveis validou agronomicamente o modelo ao isolar a inovadora métrica "Chuva na Colheita" como o preditor primordial de desastre para as culturas de inverno e soja. Adicionalmente, documentou-se o "Viés Tecnológico" na cultura do milho, limitador temporal decorrente do avanço genético não linear das sementes. Conclui-se que o framework computacional preditivo desenvolvido superou as limitações da climatologia linear padrão e dos modelos de circulação global, consolidando-se como uma ferramenta customizada e robusta para o suporte técnico ao dimensionamento de seguros agrícolas, formulação de políticas de crédito e mitigação de perdas no agronegócio paranaense.
Agriculture in the state of Paraná is characterized by high technological adoption and productivity, yet it remains structurally vulnerable to climate variability and extreme weather events. In this context, this master’s thesis aimed to develop and validate an Early Warning System (EWS) based on Machine Learning algorithms for the spatial prediction and anticipation of crop failures in the region's main rainfed crops (soybeans, maize, wheat, and dry beans). The methodology was grounded in the spatio-temporal integration of daily meteorological series from the National Institute of Meteorology (INMET) with annual agricultural yield history from the Brazilian Institute of Geography and Statistics (IBGE), spanning the 2011–2024 period. To mitigate statistical noise from continuous annual analysis, a feature engineering approach strictly coupled with local phenological calendars was proposed. Algorithmic validation avoided random splitting, instead adopting a strict Out-of-Time Validation approach, utilizing the 2011–2020 harvests for training and the climatically anomalous 2021–2024 period for blind testing. Class imbalance, inherent to natural disasters, was mitigated using the Synthetic Minority Over-sampling Technique (SMOTE) combined with Threshold Optimization.The results highlighted the "Accuracy Paradox," where complex architectures (such as MLP Neural Networks) failed to detect anomalies when evaluated by global error. By repositioning Sensitivity (Recall) as the operational success criterion, the Support Vector Machine (SVM) model with an RBF kernel excelled in the early identification of soybean crop failures, while the Random Forest algorithm showed the best trade-off for wheat. Variable importance analysis provided agronomic validation by isolating the innovative "Rain during Harvest" metric as the primary disaster predictor for winter crops and soybeans. Additionally, a "Technological Bias" was documented in maize cultivation, representing a temporal limitation due to the non-linear genetic advancement of seeds. It is concluded that the developed predictive computational framework overcame the limitations of standard linear climatology and global circulation models, establishing itself as a customized and robust tool for technical support in agricultural insurance sizing, credit policy formulation, and loss mitigation in Paraná's agribusiness.
Agriculture in the state of Paraná is characterized by high technological adoption and productivity, yet it remains structurally vulnerable to climate variability and extreme weather events. In this context, this master’s thesis aimed to develop and validate an Early Warning System (EWS) based on Machine Learning algorithms for the spatial prediction and anticipation of crop failures in the region's main rainfed crops (soybeans, maize, wheat, and dry beans). The methodology was grounded in the spatio-temporal integration of daily meteorological series from the National Institute of Meteorology (INMET) with annual agricultural yield history from the Brazilian Institute of Geography and Statistics (IBGE), spanning the 2011–2024 period. To mitigate statistical noise from continuous annual analysis, a feature engineering approach strictly coupled with local phenological calendars was proposed. Algorithmic validation avoided random splitting, instead adopting a strict Out-of-Time Validation approach, utilizing the 2011–2020 harvests for training and the climatically anomalous 2021–2024 period for blind testing. Class imbalance, inherent to natural disasters, was mitigated using the Synthetic Minority Over-sampling Technique (SMOTE) combined with Threshold Optimization.The results highlighted the "Accuracy Paradox," where complex architectures (such as MLP Neural Networks) failed to detect anomalies when evaluated by global error. By repositioning Sensitivity (Recall) as the operational success criterion, the Support Vector Machine (SVM) model with an RBF kernel excelled in the early identification of soybean crop failures, while the Random Forest algorithm showed the best trade-off for wheat. Variable importance analysis provided agronomic validation by isolating the innovative "Rain during Harvest" metric as the primary disaster predictor for winter crops and soybeans. Additionally, a "Technological Bias" was documented in maize cultivation, representing a temporal limitation due to the non-linear genetic advancement of seeds. It is concluded that the developed predictive computational framework overcame the limitations of standard linear climatology and global circulation models, establishing itself as a customized and robust tool for technical support in agricultural insurance sizing, credit policy formulation, and loss mitigation in Paraná's agribusiness.
Descrição
Citação
NAVARRO, Jorge Davi. Análise preditiva de dados geoespaciais para mitigação de desastres climáticos em propriedades rurais. 2026. Dissertação (Mestrado em Computação Aplicada) - Universidade Estadual de Ponta Grossa, Ponta Grossa, 2026.
Avaliação
Revisão
Suplementado Por
Referenciado Por
Licença Creative Commons
Exceto quando indicado de outra forma, a licença deste item é descrita como http://purl.org/coar/access_right/c_abf2
