Dry Chaco Forest deforestation map by using Random Forest with Landsat dataset on Google Earth Engine

BARRAZA, VERONICA; ROITBERG, ESTEBAN; PABLO PERNA; SALVIA, MARÍA MERCEDES; GRINGS, FRANCISCO MATIAS

Maryland

Congreso; ForestSAT2018; 2018

Semi-arid forest ecosystems play an important role in seasonal carbon cycle dynamics; however these ecosystems are prone to heavy degradation. In this research, we assess the potential to use Random Forest (RF) algorithm with the Landsat dataset on Google Earth Engine to detect cover change over the dry chaco forest (DCF), Argentina. In subtropical Argentina, the Chaco region has the highest absolute deforestation rates the country (200.000 ha/ year), and at the same time, is the least represented ecoregion in the national protected areas system. There is a critical need for methods that enable the analysis of satellite image time series to detect forest disturbances, especially in developing countries (e.g. Argentina). The unit of forest management (UMSEF) from Argentina provided an annual deforestation map based on visual inspection of Landsat images (Landsat 5 TM, Landsat 7 ETM+ and Landsat 8 OLI), taking large time of processing and the intensive and coordinated participation of many human resources. Here, we propose a RF model that automatically selects the training dataset using forest service (UMSEF) annual deforestation maps as a benchmark during 2007 to 2009. The RF training was based on 50 sample pixels for the sites in crops and another 50 over forest areas, each month from 2007 to 2009. The objective was to capture the seasonal variability of the predictor variables due to the seasonal behaviour of forest and crops. Using the monthly classification we combine the results annually with the objective of reducing false positive related to clouds, aerosols, etc. These composition were carried out using python tool development ad-hoc which assigns the most suitable class to each pixel based on measure of classification quality (CQ). These composition reduces the gaps produced by clouds and cloud shadows by assigning the correct class. The final annual product provides a nominal 30 m deforestation map over the DCF of Argentina using Landsat 7 and 8 datasets for 2007-2014. Two different metrics, derived from the confusion matrix approach (2000 independent validation samples were used per year), were selected for the accuracy assessment: the overall accuracy (OA) and the F-SCORE. In general, the global overall accuracy (OA) and F-SCORE was higher than 70% and 0.65, respectively, with a producer?s accuracy and user accuracy higher than 70%.