Predicting forest fire in the Brazilian Amazon using MODIS imagery and artificial neural networks

Abstract

The presented work describes a methodology that employs artificial neural networks (ANN) and multi-temporal imagery from the MODIS/Terra-Aqua sensors to detect areas of high risk of forest fire in the Brazilian Amazon. The hypothesis of this work is that due to characteristic land use and land cover change dynamics in the Amazon forest, forest areas likely to be burned can be separated from other land targets. A study case was carried out in three municipalities located in northern Mato Grosso State, Brazilian Amazon. Feedforward ANNs, with different architectures, were trained with a backpropagation algorithm, taking as inputs the NDVI values calculated from MODIS imagery acquired during five different periods preceding the 2005 fire season. Selected samples were extracted from areas where forest fires were detected in 2005 and from other non-burned forest and agricultural areas. These samples were used to train, validate and test the ANN. The results achieved a mean squared error of 0.07. In addition, the model was simulated for an entire municipality and its results were compared with hotspots detected by the MODIS sensor during the year. A histogram analysis showed that the spatial distribution of the areas with fire risk were consistent with the fire events observed from June to December 2005. The ANN model allowed a fast and relatively precise method to predict forest fire events in the studied area. Hence, it offers an excellent alternative for supporting forest fire prevention policies, and in assisting the assessment of burned areas, reducing the uncertainty involved in currently used methods.