Avaliação de modelos paramétricos e não paramétricos em estimativa de índice de suscetibilidade a Desertificação

• Ano de defesa: 2023
• Autor(a) principal: Santos, Thiago Costa dos
• Tipo de documento: Tese
• Tipo de acesso: Acesso embargado
• Idioma: por
• Palavras-chave em Português: Zoneamento ecológico econômico; Sensoriamento Remoto; Solo exposto; Economic ecological zoning; Remote sensing; Bare Soil; CNPQ::CIENCIAS AGRARIAS::AGRONOMIA::CIENCIA DO SOLO
• Link de acesso: http://repositorio.ufc.br/handle/riufc/75508

Desertification is process of degradation of natural resources due to the increase in anthropic pressure that leads to changes in the dynamics of ecosystems and climate variables. A significant part of the methodologies for monitoring desertification are based on models what considering index with subjective weights that increase the uncertainties of the models without identifying the causes, making it difficult to adopt efficient control strategies. This work is based on the following hypotheses: 1) parametric models (multiple regression and logistic) and non-parametric models (neural networks and SVM) are capable of predicting susceptibility to desertification, quantifying the effects of changes in vegetation, soil and climate conditions understood as causing desertification; 2) starting from the premise that the prediction models of susceptibility to desertification are restricted to local approaches without considering the precipitation regimes, it is possible that there are differences in these when used at regional and local levels and under conditions of drought or rain. Images from the TM, ETM+ and OLI sensors were used between 1997 and 2018 at the end of dry and rainy periods to quantify the NDVI, TGSI, albedo, temperature, evapotranspiration, aridity and precipitation index and to construct the geometric index of susceptibility to desertification (IGSD). The IGSD was calculated by weight the indices and constituted the Y of the models, the explanatory variables (X) were the indices themselves. The results of the models were evaluated using accuracy, kappa, average errors of omission (MEO), commission (MEC) and field hit rate (MTAC) on a local and regional scale. The results showed that the RLM and RL models produced maps with better metrics and higher MTAC at the local and regional level, on the other hand the neural network and the SVM are promising at the local level. The NDVI, albedo, evapotranspiration, temperature and aridity index were more important in the susceptibility to desertification at the local and regional level, and the increase in the NDVI and aridity index reduced the IGSD average and the risk of desertification, while the increases in albedo, temperature and evapotranspiration resulted in increased in IGSD averages and in risks to desertification. Disorderly deforestation, increase in areas occupied by pastures and degraded caused the reduction of vegetation cover and increase in albedo, temperature and evapotranspiration, impacting precipitation and the aridity index, increasing the risks of desertification. Therefore, parametric and non-parametric models have the potential to identify susceptibility to desertification and its causes, in addition to showing in quantitative terms the contribution of each index in increasing or reducing the risk of desertification.