Séries temporais TM/RapidEye para a detecção de pequenos distúrbios em florestas tropicais

Abstract

This dissertation was developed based on the need for new methodologies to assist in the inspection of remaining native vegetation, especially for areas with a high level of forest fragmentation. Furthermore, considering the tropical ecosystems, they are essential for the maintenance of ecological services, and shelter the world's largest biodiversity. However, deforestation represents one of the principal human activities that have affected the stability of these environments. Thus, monitoring the dynamics of tropical forests is critical for decision-making for public and environmental management of these ecosystems. Given this context, this work was structured in two chapters, in which the first chapter covers all problematic for monitoring of small-scale deforestation in tropical ecosystems, which have a great anthropic interference. The second chapter includes the application of a new methodology for the detection of small deforestations, considering the knowledge gaps for monitoring the vegetation in tropical ecosystems. Thus, the aim of this study was to propose a new methodology using object-based multi-sensor time series (TM/RapidEye), to reduce the computational time and improve the accuracy of automatic detection of small deforestations, applying the BFAST Monitor algorithm. The study was developed in an Atlantic Forest area, located in Santa Catarina, in southern Brazil. In this biome, the deforestation process is characterized by small punctual events dispersed in an anthropic matrix. For the construction of the time series, 230 TM images were acquired (path/row 227/068), between 1984 and 2011, and 20 RapidEye images (tile 2226122) from 2009 to 2011. Eight time series were built with object-images, the value of each pixel corresponding to the value of the statistical parameter extracted from each object (average or minimum) in each image of the time series. In the analysis by BFAST Monitor it is necessary to define a historical period and a monitoring period. A regression model is then fit to the data set identified as the stable historical period. The identification of this period is critical f or distinguishing between natural and abrupt changes (changes of interest). Subsequently, stability in this model is tested for the monitoring period data set. If the structural stability hypothesis is rejected, a breakpoint (potential change) is detected. In this work, the historical period was set between 1984 and 2010 and the monitoring period as the year 2011. The results elucidated great potential on the methodology, since the use of object-images reduced processing time by 95%, and inserting RapidEye images for building multi-sensor time series, provided a fourfold increase in the automatic detection capability of small deforestations identified in the assessed region.