2023-10-06, 10:00–10:30, EURAC Auditorium
This talk will discuss how Brazil's National Institute for Space Research (INPE) is transitioning from visual interpretation to automatic classification of its Amazon deforestation monitoring system. The presentation will discuss the methods for machine learning using time series of Sentinel-2/2A images that have managed to reach the same accuracy as remote sensing experts.
The Brazilian National Institute for Space Research (INPE) makes yearly assessments of deforestation in the Brazilian Amazonia rain forest since 1988, INPE’s data is considered authoritative by the scientific community, national and international media, and is used as a basis for Brazil’s reporting of greenhouse gases (GHG) emissions. Due to the need for high accuracy of the result, the PRODES maps are produced by visual interpretation and have a 93% accuracy. However, doing visual interpretation of an area of 4 million km2 requires substantial human resources. For this reason, INPE is working to transition the current data production technique to a method based on big data analytics using Sentinel-2 and Sentinel-2A images.
The INPE team has developed new methods for producing deforestation and land use and land cover change maps using satellite image time series. The method requires a regular data cube of dense satellite images. Experts provide training samples that represent the possible deforestation events. These samples train machine learning and deep learning classifiers, which work on a time-first, space-later basis. First, each pixel location is associated to a time series that is classified. Then, a neighborhood-based Bayesian smoother is applied to remove outliers and produce results that approximate those of visual interpreters. The resulting maps have a 98% coincidence with those done by human experts.
The presentation will discuss the method and its implementation in an open-source software R package. The package is one of the components of the Open Earth Monitor project.
OEMC Grant agreement ID: 101059548
Researcher on Geoinformatics, Land Use Change and Spatial Data Science. Former Director of INPE/Brazil (2005-12) and of Group on Earth Observations (2018-21).