| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 13.22 MB | Adobe PDF |
Authors
Abstract(s)
A agricultura em Portugal, especialmente nas áreas de montado, enfrenta sérios desafios devido
às alterações climáticas, que comprometem a produtividade das pastagens. Contudo, a produtividade
das pastagens não é apenas afetada pelo clima, mas também por diversos fatores ambientais locais. A
utilização de técnicas de deteção remota e modelação revelou-se uma ferramenta importante para avaliar
os padrões espácio-temporais de produtividade. Através de índices de vegetação como o NDVI e o PPI,
calculados a partir de imagens multiespectrais, é possível monitorizar a produtividade e identificar áreas
com maior risco de degradação. Estes indicadores, quando combinados com variáveis bioclimáticas e
características do terreno, permitem modelar a variabilidade espacial das pastagens. No Alentejo, na
região de Mértola, foram selecionadas parcelas de estudo onde se aplicou o algoritmo de aprendizagem
automática XGBoost para avaliar a produtividade das pastagens. O modelo correlacionou índices de
vegetação de alta resolução com variáveis ambientais. Foram geradas previsões de produtividade para
os anos de 2030, 2050, 2070 e 2100, considerando diferentes cenários climáticos (RCP 2.6, 4.5 e 8.5).
Este estudo insere-se no projeto AdaptForGrazing, financiado pelo Plano de Recuperação e Resiliência,
PRR, com o objetivo de identificar os principais fatores locais que influenciam a produtividade das
pastagens, definir metodologias para modelar esta produtividade e testar a aplicação do XGBoost para
dados espácio-temporais. Concluiu-se que os fatores locais mais determinantes são as práticas de gestão,
a elevação e o índice topográfico de humidade (TWI), que apontam para zonas mais baixas com maior
acumulação de água e maior produtividade e ainda temperatura e precipitação de inverno. O modelo
preditivo para cenários futuros revelou um erro elevado e um R² baixo (~0.3), devido à limitada
capacidade de extrapolação do XGBoost. A previsão da produtividade futura de pastagens exige dados
com maior variabilidade climática que os disponíveis na área de estudo.
Agriculture in Portugal, particularly in montado regions, faces significant challenges due to climate change, which compromises pasture productivity. However, local environmental factors also play a critical role in influencing this productivity beyond just climate. Remote sensing and modeling techniques offer valuable tools to assess the spatio-temporal patterns of pasture productivity. Vegetation indices such as NDVI and PPI, derived from multispectral remote sensing imagery, enable the monitoring of productivity and the identification of more or less degraded areas. These indicators can be combined with bioclimatic variables and terrain characteristics to model spatial variability in pastures. In the Alentejo region, near Mértola, study plots were selected to apply the XGBoost machine learning algorithm to evaluate pasture productivity. This was done to evaluate pasture productivity by correlating high-resolution vegetation indices with environmental variables. Predictive indicators of productivity were also calculated using an XGBoost model for the years 2030, 2050, 2070, and 2100, considering various climate scenarios (RCP 2.6, 4.5, and 8.5). This research is part of the AdaptForGrazing project, funded by the Recovery and Resilience Plan (PRR), with the goals of identifying the local factors that determine pasture productivity, establishing a workflow for modeling productivity at the property scale, and testing XGBoost’s ability to model pastureland based on spatiotemporal data. It was concluded that the main local factors determining pasture productivity in the study area are management practices, elevation, and the Topographic Wetness Index (TWI). Lower areas with higher water accumulation showed greater productivity. Winter temperatures and precipitation also played an influential role. The predictive model for future climate scenarios exhibited high error rates and low R² (~0.3), due to XGBoost’s limited extrapolation capacity. Predicting future pasture productivity requires data with greater climatic variability than currently available in the study area.
Agriculture in Portugal, particularly in montado regions, faces significant challenges due to climate change, which compromises pasture productivity. However, local environmental factors also play a critical role in influencing this productivity beyond just climate. Remote sensing and modeling techniques offer valuable tools to assess the spatio-temporal patterns of pasture productivity. Vegetation indices such as NDVI and PPI, derived from multispectral remote sensing imagery, enable the monitoring of productivity and the identification of more or less degraded areas. These indicators can be combined with bioclimatic variables and terrain characteristics to model spatial variability in pastures. In the Alentejo region, near Mértola, study plots were selected to apply the XGBoost machine learning algorithm to evaluate pasture productivity. This was done to evaluate pasture productivity by correlating high-resolution vegetation indices with environmental variables. Predictive indicators of productivity were also calculated using an XGBoost model for the years 2030, 2050, 2070, and 2100, considering various climate scenarios (RCP 2.6, 4.5, and 8.5). This research is part of the AdaptForGrazing project, funded by the Recovery and Resilience Plan (PRR), with the goals of identifying the local factors that determine pasture productivity, establishing a workflow for modeling productivity at the property scale, and testing XGBoost’s ability to model pastureland based on spatiotemporal data. It was concluded that the main local factors determining pasture productivity in the study area are management practices, elevation, and the Topographic Wetness Index (TWI). Lower areas with higher water accumulation showed greater productivity. Winter temperatures and precipitation also played an influential role. The predictive model for future climate scenarios exhibited high error rates and low R² (~0.3), due to XGBoost’s limited extrapolation capacity. Predicting future pasture productivity requires data with greater climatic variability than currently available in the study area.
Description
Tese de Mestrado, Engenharia da Energia e do Ambiente, 2024, Universidade de Lisboa, Faculdade de Ciências
Keywords
Índices de vegetação PPI XGBoost Alentejo Alterações climáticas Teses de mestrado - 2024