Desarrollo de una metodología para la detección de cambios de la cubierta vegetal en series temporales de imágenes de satélite diarias. Aplicación a la detección de áreas quemadas

We have developed a methodology for detection of observable phenomena at pixel level over time series of daily satellite images, based on using a Bayesian classifier. This methodology has been applied successfully to detect burned areas in the North American boreal forests using the LTDR dataset. The LTDR dataset represents the longest time series of global daily satellite images with 0.05° (~5 km) of spatial resolution. The proposed methodology has several stages: 1) pre-processing daily images to obtain composite images of n days; 2) building of space of statistical variables or attributes to consider; 3) designing an algorithm, by selecting and filtering the training cases; 4) obtaining probability maps related to the considered thematic classes; 5) post-processing to improve the results obtained by applying multiple techniques (filters, ranges, spatial coherence, etc.). The generated results are analyzed using accuracy metrics derived from the error matrix (commission and omission errors, percentage of estimation) and using scattering plots against reference data (correlation coefficient and slope of the regression line). The quality of the results obtained improves, in terms of spatial and timing accuracy, to other burned area products that use images of higher spatial resolution (500 m and 1 km), but they are only available after year 2000 as MCD45A1 and BA GEOLAND-2: the total burned area estimation for the study region for the years 2001-2011 was 28.56 millions of ha according to reference data and 12.41, 138.43 and 19.41 millions of ha for the MCD45A1, BA GEOLAND-2 and BA-LTDR burned area products, respectively