Article Preview
TopIntroduction
Forest fires are among the worst natural catastrophes because of their fast spread and lack of controllability. As a result, forest fire prevention has become a major concern in the disciplines of forestry and ecology. Algeria is a country located on the southern rim of the Mediterranean Basin. It is highly affected by forest fires and each year millions of forest hectares (ha) are destroyed. Forest fires are usually due to a climate favorable to ignition, propagation and the abundance of combustible materials such as shrublands and woods (Curt et al., 2020). The 2020 fire season was especially dramatic, hectares were burnt by 3,493 hot spots, of which 38 percent were forests (representing 16,570 hectares), 32 percent were bushes, and 30 percent were vegetation. Algeria is the fourth most impacted country among those covered by the European Forest Fire Information System (EFFIS) (San-Miguel-Ayanz et al., 2017).
Early fire prediction and detection are critical steps that will greatly decrease the disaster's damage and firefighting efforts. Many techniques have been developed, including approaches that use satellite images, historical weather data and computational fluid dynamics. Although many machine learning models (ML) have been used to predict forest fires (Abid, 2021), to the best of the authors’ knowledge, only a few references in the literature systematically describe the effect of using these models on embedded devices. This was the motivation behind the present study.
Neural Network models that use meteorological data are known to influence forest fires. These data are detected by wireless sensor networks (WSNs). Therefore, they may be collected in real-time at very low costs compared to satellite and scanner data. The adoption of WSNs data can help to eliminate false positives produced by transient changes in a single sensor's output. Furthermore, the multiple data fusion technique reduces communication costs while also conserving energy.
This article makes the following key contributions:
- 1.
The researchers present an overview of state-of-the-art fire prediction techniques based on WSN sensors and machine learning.
- 2.
Using two WSN datasets, they investigate and compare the performance of nine Machine Learning algorithms.
- 3.
The models are compared and evaluated based on Accuracy, F-score, AUC-ROC, memory consumption, training and prediction times.
- 4.
The authors propose an embedded forest fire prediction model based on ANN using the results obtained from comparisons.
The following section discusses the related work. Section 3 primarily covers the proposed work. Section 4 presents the results of the performance evaluation of classification models and their comparative analysis. Finally, Section 5 summarises the conclusion and future scope of this work.
TopIn this study, the authors investigated at several approaches proposed for the forest fire prediction. There have been several works in the literature conducted using machine learning models (Abid, 2021). The ML models that have been adopted in this context are Artificial Neural Networks (ANN) (Yan et al.,2016; Liu et al., 2011;Yu et al., 2005; Hefeeda & Bagheri, 2007), logistic regression (Chang et al., 2013; Catry et al., 2009; Chuvieco et al., 2009; Kalabokidis et al., 2002; De Vasconcelos et al., 2001), decision tree (DT) and trees-based models (Giuntini et al., 2017; Pourtaghi et al., 2016; Maksimović et al., 2013; Oliveira et al., 2012; Stojanova et al.,2010; Lozano et al.,2008; Prasad et al., 2006).