Home // SENSORCOMM 2011, The Fifth International Conference on Sensor Technologies and Applications // View article

Machine Learning and Dataming Algorithms for Predicting Accidental Small Forest Fires

Authors:
Vasanth Iyer
S. Sitharama Iyengar
Paramesh Nandan
Rama Murthy Garmiela
M.B. Srinivas Mandalika

Keywords: Machine Learning, Datamining, Naive Bayes, Forest fires, FWI, Temporal Patterns, WEKA machine learning framework.

Abstract:
Extracting useful temporal and spatial patterns from sensor data has been seen before, the technical basis of Machine learning with Data mining is studied with the evidence collected uniformly over many years and which allow using users' perspective in collected evidence. This model helps in probabilistically forecasting fires and help forest department in planing day to day schedules. Using a model to predict future events reliably one needs to collect samples from sensors and select a feature, which does have any particular bias. Due to practicable problems most of the collected data have 80% of attributes missing and the remaining has numeric values, which are hard to discretization. To adapt to such limitations, we use nominal data type, which allows better understanding of the temporal and spatial features, which are learnt. We encounter several practicable limitations as forest fires events are very rare and manual classification is extremely costly. Another is the unbalanced nature of the problem of the many forest fire events many are of the burnt area is very small and gives skewed distribution. Most of the examples naturally group into batches, which are collected from evidence satellite photography and collaborative reports from national parks departments. The second set of database was collected from the meteorological weather station about several weather observations, which are located very close to the reported fires. Finally, the compiling task is to serve as a filter and provide the user to vary the false alarm rate. We show by regression analysis of the compiled dataset that the forest fire classifier has a minimum false alarm rate when including temporal features. The machine learning algorithms successfully classifies accidental small fires with 85% reliably and large fires by a much lower accuracy of 30%.

Pages: 116 to 121

Copyright: Copyright (c) IARIA, 2011

Publication date: August 21, 2011

Published in: conference

ISSN: 2308-4405

ISBN: 978-1-61208-144-1

Location: Nice/Saint Laurent du Var, France

Dates: from August 21, 2011 to August 27, 2011