Burn Area Processing to Generate False Alarm Data for Hotspot Prediction Models
Abstract: Developing hotspot
prediction models using decision tree algorithms require target classes to
which objects in a dataset are classified.
In modeling hotspots occurrence, target classes are the true class
representing hotspots occurrence and the false class indicating non hotspots
occurrence. This paper presents the
results of satellite image processing in order to determine the radius of a
hotspot such that random points are generated outside a hotspot buffer as false
alarm data. Clustering and majority
filtering were performed on the Landsat TM image to extract burn scars in the
study area i.e. Rokan Hilir, Riau Province Indonesia. Calculation on burn areas and FIRMS MODIS
fire/hotspots in 2006 results the radius of a hotspot 0.90737 km. Therefore, non-hotspots were randomly
generated in areas that are located 0.90737 km away from a hotspot. Three
decision tree algorithms i.e. ID3, C4.5 and extended spatial ID3 have been
applied on a dataset containing 235 objects that have the true class and 326
objects that have the false class. The results are decision trees for modeling
hotspots occurrence which have the accuracy of 49.02% for the ID3 decision
tree, 65.24% for the C4.5 decision tree, and 71.66% for the extended spatial
ID3 decision tree.
Author: Imas S Sitanggang,
Razali Yaakob, Norwati Mustapha, Ainuddin A. N
Journal Code: jptkomputergg150108
