Análise de desempenho de produtos modis para modelagem da dinâmica de inundação do Pantanal mato-grossense

• Autor(a) principal: Silveira, Sérgio Wagner Gripp da
• Data de Publicação: 2015
• Tipo de documento: Tese
• Idioma: por
• Título da fonte: Repositório Institucional da UFMT
• Texto Completo: http://ri.ufmt.br/handle/1/2183
• topic: Regressão logística; MOD13Q1; LSWI; EVI; NDVI; CNPQ::CIENCIAS EXATAS E DA TERRA::FISICA; Logistic regression; MOD13Q1; LSWI; EVI; NDVI

The ecological processes of wetlands, such as the Pantanal of Mato Grosso, depend on their cyclic flooding, which field monitoring is unfeasible due to its vast extension and difficult access. Thus this study aimed to evaluate the applicability of MODIS products to monitor the inundation dynamics, assessing systematically the interference of field sampling scheme, pixel quality given by the Quality Assurance band, vegetation cover, period of the hydrological cycle and the use of a gap filling algorithm on the classification model performance. For logistic regression classification training and validation, inundation measurements of two field survey sites were considered: The Pirizal site of the Long Term Ecological Research Program (LTER/PELD), and the Private Natural Heritage Preservation Area (RPPN) of the SESC-Pantanal. Binary coded flood measurements were compared with the NDVI, EVI and LSWI spectral indices obtained from the MOD13Q1 product and annual land use and cover maps of the MCD12Q1 product. Degree of generalizability of the model was measured by external and cross-validation. A new method, called virtual stages, was introduced to test the influence of field point measurements with zonal entities represented by MODIS pixels. In this procedure, flooding levels in regular 25m grid around the limnimetric gauges were simulated by a Digital Elevation Model (DEM) interpolated from an existing topographic survey. The combination of the LSWI, EVI and land use and cover layer resulted in the best flooding predictions with a Nagelkerke R2 = 0.75 and Kappa = 0.78. The overall accuracy was 89.2% for the adjusted model and 77.2% in the external validation. A model setup using the virtual gauges did not result in a better overall performance of the models. Nevertheless, false positive rates were partially reduced. The omission of pixels down to a QA level of 0110 did not significantly improve the model performance. An expressive performance loss was only observed when pixels flagged as cloudy were included in the models. The filtered and filled in model showed excellent results, if compared with models based on different QA filter levels, having as further advantages the gap-free flood monitoring. Model performance was reduced in forested areas, however less expressive than expected, if compared with the performance in open savannah formations. Model performance varied seasonally, being reduced during rising and falling waters. Differences higher than 10% in the performance indicators showed that only an external validation allow a realistic assessment of flood mapping performance with MODIS data in the floodplain.