A carregar...
Publicação
Aplicação de previsão edafoclimática e aptidão cultural em Portugal Continental
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 2.06 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
A agricultura em Portugal enfrenta elevada incerteza climática e grande variabilidade de solos.Esta incerteza tem implicações diretas na decisão sobre as culturas mais adequadas a cada região, tornando essencial dispor de ferramentas que orientem o agricultor com base em dados objetivos. Para apoiar decisões práticas sobre que cultura plantar em cada local foi criado um sistema de apoio à decisão que prevê variáveis edafoclimáticas e converte essas previsões em rankings de aptidão com indicação dos fatores limitantes. O trabalho cobre todo o ciclo de ciência de dados desde a recolha e integração de informação de 158 localizações até à modelação e à disponibilização numa interface interativa. O modelo preditivo baseia-se em Multi-Output Random Forest e estima, em simultâneo, horas de frio, dias de geada, temperaturas, precipitação e atributos do solo. A interface permite escolher uma localização em Portugal continental e obter a aptidão de oito culturas mediterrânicas com visualizações claras e recomendações acionáveis.Os resultados mostram desempenho elevado, com coeficiente de determinação global próximo de 0,95 e especial robustez nas variáveis climáticas. As variáveis de solo apresentam menor precisão devido à maior variabilidade local, sem comprometer a utilidade do sistema. A comparação com catorze modelos, desenvolvidos individualmente para cada varíavel climática ou de solo, indica ligeiro ganho de exatidão pontual, mas menor capacidade de generalização, favorecendo a abordagem multioutput para utilização real. Como medidas de melhoria do sistema desenvolvido, recomenda-se ampliar o leque de culturas e incluir variedades, enriquecer a base de dados e as covariáveis, criar uma base permanente com monitorização contínua, validar em mais regiões do país e integrar cenários futuros através de modelos mecanísticos, fortalecendo a resiliência das decisões agrícolas.
Agriculture in Portugal faces high climatic uncertainty and significant soil variability. This uncertainty directly affects decisions regarding which crops are most suitable for each region, highlighting the need for tools that can guide farmers through data-driven insights. To support practical decisions on which crop to plant in each location, a decision support system was developed that predicts edaphoclimatic variables and converts these predictions into suitability rankings, identifying the main limiting factors. The work covers the entire data science cycle—from the collection and integration of information from 158 locations to modelling and the development of an interactive interface. The predictive model is based on a Multi-Output Random Forest, capable of simultaneously estimating chilling hours, frost days, temperature, precipitation, and soil attributes. The interactive interface allows users to select any location in mainland Portugal and obtain crop suitability results for eight Mediterranean species, presented through clear visualisations and actionable recommendations. The results show high predictive performance, with an overall coefficient of determination close to 0.95 and particularly strong robustness for climatic variables. Soil-related variables display lower precision due to greater local variability, without compromising the system’s overall usefulness. A comparison with fourteen single-output models created for each climatic and soil variables indicates slightly higher point accuracy but lower generalisation capacity, favouring the multi-output approach for real-world application. As an improvement of the current system, it is recommended to expand the range of crops and include intra-species varieties, enrich the database and covariates, establish a permanent and continuously monitored data infrastructure, validate the model in additional regions of the country, and integrate future climate scenarios through mechanistic modelling—thereby strengthening the resilience of agricultural decision-making.
Agriculture in Portugal faces high climatic uncertainty and significant soil variability. This uncertainty directly affects decisions regarding which crops are most suitable for each region, highlighting the need for tools that can guide farmers through data-driven insights. To support practical decisions on which crop to plant in each location, a decision support system was developed that predicts edaphoclimatic variables and converts these predictions into suitability rankings, identifying the main limiting factors. The work covers the entire data science cycle—from the collection and integration of information from 158 locations to modelling and the development of an interactive interface. The predictive model is based on a Multi-Output Random Forest, capable of simultaneously estimating chilling hours, frost days, temperature, precipitation, and soil attributes. The interactive interface allows users to select any location in mainland Portugal and obtain crop suitability results for eight Mediterranean species, presented through clear visualisations and actionable recommendations. The results show high predictive performance, with an overall coefficient of determination close to 0.95 and particularly strong robustness for climatic variables. Soil-related variables display lower precision due to greater local variability, without compromising the system’s overall usefulness. A comparison with fourteen single-output models created for each climatic and soil variables indicates slightly higher point accuracy but lower generalisation capacity, favouring the multi-output approach for real-world application. As an improvement of the current system, it is recommended to expand the range of crops and include intra-species varieties, enrich the database and covariates, establish a permanent and continuously monitored data infrastructure, validate the model in additional regions of the country, and integrate future climate scenarios through mechanistic modelling—thereby strengthening the resilience of agricultural decision-making.
Descrição
Dissertação de Mestrado, Ciência de dados em Agricultura, Alimentação, Floresta e Ambiente, 2025, Universidade de Lisboa, Instituto Superior de Agronomia.
Palavras-chave
agricultura orientada por dados ciência de dados sistemas de apoio à decisão machine learning aptidão agrícola data-driven agriculture data science decision support systems machine learning crop suitability