Estimating Forest Losses Using Spatio-temporal Pattern-based Sequence Classification Approach

Abstract

Consistent forest loss estimates are important to enforce forest management regulations. In Tunisia, recent evidence has suggested that the deforestation rate is increasing, especially since the 2011’s Revolution. However, no spatially explicit data on the extent of deforestation before and after the Revolution exists. Here, we quantify deforestation in the country for the period 2001–2014 and we propose a novel spatio-temporal pattern-based sequence classification framework for forest loss estimation. To do so, expert knowledge and spatial techniques are applied to identify deforestation drivers. Then, we adopt sequential pattern mining to extract sets of patterns sharing similar spatiotemporal behavior. The sequence miner generates multidimensional-closed sequential patterns at different time granularities. Then, a discriminative filter is employed to decide on patterns to use as relevant classification features. Lastly, the classifier is trained using random forest and shows an improved result.