ForestEyes - using citizens to track deforestation

Conservation of tropical forests is a relevant issue because of their important role in the global ecosystem. They have a great diversity of fauna and flora, besides regulating the climate and rainfall, absorbing large quantities of carbon dioxide and being indigenous dwellings. Unfortunately, millions of hectares have been lost and degraded over the years.

The rainforests need to be protected and collective action to complement existing initiatives of conservation need to be created.

The ForestWatchers Project was created in 2012 with this purpose: to ally Citizen Science with forest monitoring. In one of its applications, remote sensing images of forested areas were classified into forest or non-forest with an automated classification algorithm and the accuracy of the resulting map could be further improved by volunteer observation on the Web.

Inspired by this application of ForestWatchers, this Citizen Science project wants citizens to help us to track deforestation. Volunteers will be classifying latest remote sensing tiles that are going to be the training set of a classification algorithm. This procedure is expected to be fast and cheap, helping to track deforestation even in places where does not exist any official monitoring programs on tropical forests.

To determine which remote sensing tiles are going to be analyzed by the volunteers, techniques of a machine learning paradigm known as Active Learning will be used. It has the goal to incrementally build a training set with a small number of samples but with high representativeness, achieving maximum accuracy in the classification of new samples. For this application, the procedure will be: we start with a very small training set to automatic classify remote sensing images' pixels in forest or non-forest; the active learning technique will choose new potential samples to be included in the training set; these samples will be analyzed and classified by volunteers in this citizen science project; the samples classified here will be added to the training set and the automatic classifier will be trained and the active learning technique will choose new samples to be analyzed and classified by the volunteers and so on. This procedure will be carried out until the automatic classifier reachs a reasonable hit rate.

The images currently used in this project are images obtained by NASA's MODIS sensor (https://modis.gsfc.nasa.gov/) and by Landsat-8/OLI satellite (https://landsat.usgs.gov/landsat-8). The MODIS images were kindly provided by the ForestWatchers Project team and the images from Landsat-8/OLI are available for free by USGS (United States Geological Survey) at https://earthexplorer.usgs.gov/. In the future images from other satellites may be used.

For preliminary studies and validation of this project were carried out tasks with the same data used in ForestWatchers Project: images from MODIS sensor for the areas of Rondônia state (Brazil) in the year 2011 and the indigenous reservation of Awá-Guajá (Maranhão state - Brazil) in the year 2014. Currently tasks are being carried out with Landsat-8/OLI images of Rondônia state in the year 2016. The obtained results for these and for future tasks will always be available in Results.

The current application is LANDSAT 8 - Segments. With a method called Simple Linear Iterative Clustering (SLIC) a remote sensing image from an area of Rondônia State (Brazil) was segmented in approximately 1000 pieces. We need the volunteers to classify each segment in Forest, Non-Forest or Undefined. Each segment can be seen in a RGB composition image or a false-color composition with bands 7, 5 and 3 of LANDSAT-8.

The classified segments will be used to train an automatic classifier that will analyze new remote sensing imagery to track rainforests' deforestation.