Metodologia para detecção de colheita de eucalyptus no Espírito Santo: uma abordagem com sentinel-2

• Autor(a) principal: Marques, Leon Muller
• Data de Publicação: 2023
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da Universidade Federal do Espírito Santo (riUfes)
• Texto Completo: http://repositorio.ufes.br/handle/10/12497
• topic: Recursos Florestais e Engenharia Florestal; Eucalipto; Sensoriamento remoto; Cobertura do solo; Floresta plantada

The public sector requires tools enabling systematic monitoring of cultivated forest areas. The use of remote sensing techniques facilitates large-scale surveillance by environmental agencies. The primary aim of this study was to create a methodology for identifying Eucalyptus stands' harvesting in the state of Espírito Santo. To conduct this monitoring, the Eucalyptus stands' database provided by a company was utilized. This database was constructed from visual analysis of Sentinel-2 images from the year 2020. The accuracy of the map was validated using the Kappa coefficient. Reference data for harvesting was also obtained through visual interpretation by analyzing a fiveyear history of Sentinel-2 images. Each year, the best images from each quarter were selected, identifying changes in spectral response from forest to exposed soil, indicating harvesting.In constructing the harvest detection algorithm, land cover classes provided by the European Space Agency (ESA) between 2019 and 2020 were used, employing the Scene Classification (SCL). This product offers land cover information for each Sentinel-2 image collected every five days. Subsequently, classes of interest (cloud, soil, and vegetation) were filtered. To acquire an image with fewer cloud cover instances, three temporal categories (monthly, bimonthly, and quarterly) were analyzed, generated by aggregating weekly images. SCL classes were intersected with Eucalyptus stands, determining the temporal category with the least cloud cover in the stand database.Once the best temporal category was defined, harvest detection relied on accumulating pixels classified as soil in each new image composition, evaluating the percentage of soil in the stand. Experimentations were conducted to define the best soil percentage threshold to consider a stand as harvested. To calculate algorithm accuracy, performance evaluations were done with two distinct strategies. Strategy 1 directly compared the algorithm-identified harvest month with the reference date for each Eucalyptus stand. Strategy 2 assessed algorithm accuracy by varying one month before and after the reference date. Subsequently, the impact of slope and stand size on algorithm accuracy was analyzed, calculating errors and successes for each slope class and stand size.The Eucalyptus map obtained a Kappa of 0.851 and an overall accuracy of 94%. The quarterly temporal category proved most effective in minimizing cloud effects, as no stand exhibited over 20% cloud coverage. Strategy 2 was the most efficient, achieving an algorithm accuracy of 84.5% with a 25% soil threshold. It was observed that higher slopes corresponded to lower accuracy, while stand size showed a direct relationship: larger size led to higher accuracy.The developed algorithm represents an advancement in applying new monitoring and oversight methods for Eucalyptus plantations, and it can be adopted by the public sector.