Jesús A. Anaya

Accuracy Assessment of Burned Area Products in the Orinoco Basin

Burned area products derived from satellite images are used as input to determine biomass burning emissions. Appropriate assessment of the accuracy of burned area products is required to assess reliable emissions. This document provides validation results for four burned area products: GlobCarbon, MCD45, L3JRC and AQS. The study area is at the northern South American savannas along the Orinoco River since there is a rapid conversion of Amazonian forest to cattle pasture. A validation method was applied from 2001 to 2007 based on the comparison of commission and omission errors from 20 confusion matrixes with their respective efficient solution. Efficient solutions were determined using the “Pareto Boundary”. This method allows estimating the potential for improving burned area algorithms as well as evaluating the effect of pixel size on accuracy. A landscape metric was used to analyze the weight of the fragments’ distribution on global accuracy. It was found that all products underestimate burned area and that an increase in pixel size or border density results in larger burned area estimate errors.

Land cover mapping of a tropical region by integrating multi-year data into an annual time series

Generating annual land cover maps in the tropics based on optical data is challenging because of the large amount of invalid observations resulting from the presence of clouds and haze or high moisture content in the atmosphere. This study proposes a strategy to build an annual time series from multi-year data to fill data gaps. The approach was tested using the Moderate Resolution Imaging Spectroradiometer (MODIS) vegetation index and spectral bands as input for land cover classification of Colombia. In a second step, selected ancillary variables, such as elevation, L-band Radar, and precipitation were added to improve overall accuracy. Decision-tree classification was used for assigning eleven land cover classes using the International Geosphere-Biosphere Programme (IGBP) legend. Maps were assessed by their spatial confidence derived from the decision tree approach and conventional accuracy measures using reference data and statistics based on the error matrix. The multi-year data integration approach drastically decreased the area covered by invalid pixels. Overall accuracy of land cover maps significantly increased from 58.36% using only optical time series of 2011 filtered for low quality observations, to 68.79% when using data for 2011 ± 2 years. Adding elevation to the feature set resulted in 70.50% accuracy.

Integrating remotely sensed fires for predicting deforestation for REDD

Fire is an important tool in tropical forest management, as it alters forest composition, structure, and the carbon budget. The United Nations program on Reducing Emissions from Deforestation and Forest Degradation (REDD+) aims to sustainably manage forests, as well as to conserve and enhance their carbon stocks. Despite the crucial role of fire management, decision-making on REDD+ interventions fails to systematically include fires. Here, we address this critical knowledge gap in two ways. First, we review REDD+ projects and programs to assess the inclusion of fires in monitoring, reporting, and verification (MRV) systems. Second, we model the relationship between fire and forest for a pilot site in Colombia using near-real-time (NRT) fire monitoring data derived from the Moderate Resolution Imaging Spectroradiometer (MODIS). The literature review revealed fire remains to be incorporated as a key component of MRV systems. Spatially explicit modeling of land use change showed the probability of deforestation declined sharply with increasing distance to the nearest fire the preceding year (multi-year model area under the curve [AUC] 0.82). Deforestation predictions based on the model performed better than the official REDD early-warning system. The model AUC for 2013 and 2014 was 0.81, compared to 0.52 for the early-warning system in 2013 and 0.68 in 2014. This demonstrates NRT fire monitoring is a powerful tool to predict sites of forest deforestation. Applying new, publicly available, and open-access NRT fire data should be an essential element of early-warning systems to detect and prevent deforestation. Our results provide tools for improving both the current MRV systems, and the deforestation early-warning system in Colombia.