Analysis of the influence of spatial variables on the location of deforestation in the Brazilian Amazon
Britaldo
Silveira
Soares Filho, Universidade Federal de Minas Gerais, britaldo@csr.ufmg.br
(Presenting)
Hermann
O.
Rodrigues, Universidade Federal de Minas Gerais, hermann@csr.ufmg.br
Daniel
Curtis
Nepstad, Woods Hole Research Center, dnepstad@whrc.org
Gustavo
Coutinho
Cerqueira, Universidade Federal de Minas Gerais, cerca@csr.ufmg.br
Eliane
Voll, Universidade Federal de Minas Gerais, voll@csr.ufmg.br
Ane
Alencar, Instituto de Pesquisa Ambiental da Amazônia, ane@ipam.org.br
Spatially explicit simulations of deforestation rely on the calculation of probability maps, which attempt to quantify and integrate the influences of variables, representing biophysical, infrastructure, and territorial features - such as topography, rivers, vegetation, soils, climate, proximity to roads, towns and markets, and land use zoning -, on the spatial prediction of deforestation. Previous analytical modeling of deforestation included mainly regression methods like logistic regression or weights of evidence. We have developed a heuristic method of analyzing the effects of spatial variables on the location of deforestation by applying genetic algorithm (GA) to calculate probability surfaces of deforestation. The GA takes advantage of the weights of evidence method using its same formulas but that are now calibrated through the GA selection mechanisms. The developed method was tested in 12 case study regions representative of different types of Amazonian colonization frontier, each one comprising a Landsat scene. Database for the selected regions includes INPE/PRODES deforestation maps from 1997 to 2000, at 250 meter resolution, and cartographic layers of road and urban networks, soils, vegetation, topography, rivers, settlement and protected areas, and distance to previously deforested land. The results from GA method were assessed comparing the simulated 1997-2000’s deforestation map with PRODES 2000’s map through image similarity test based on a fuzzy multiple resolution comparison. GA showed better performance than the weights of evidence method, achieving agreements up to 44% at cell by cell comparison and up to 83% for a window size 5x5. This analysis also pointed out the variables “distance to roads” and “distance to previously deforested land” to be the strongest regional predictors of deforestation.
Submitted by Britaldo Soares Filho on 18-MAR-2004
Science Theme: LC (Land Use and Land Cover Change)