A machine learning algorithm for predicting wild fire occurrence
Otieno, Jack Odunga
MetadataShow full item record
A wild fire is an unplanned fire that burns in a natural area such as forest and grassland. Indeed, wild fires are destroying parts of the world’s forest coverage, affecting land and killing wild life. Majority of wild fires are caused by human activities and weather conditions. Despite various forest management authorities using numerous methods to detect and suppress fire, there are several wild fires reported around the globe annually. In Kenya for instance, Kenya Forest Services ensures fire detection and detention through the use of ground patrols and fixed stations (fire towers). They also use radio systems, vehicles, motorcycles and even bicycles. These techniques are not working effectively due to inadequate staff and resources to cover the available forest coverage. The research proposed the development of supervised machine learning model to predict wild fires using existing data that was collected from credible climatological sources that included both meteorological sources as well as wildfire databases focusing on content dated from 2000 to 2020. The study utilized data sets from multiple sources including National Fire Danger Rating System (NFDRS), Canada National Fire Database (CNFDB), University of California machine learning repository, and scientifically verified Internet sources. The methodology involved collection of relevant data sets, cleaning and preparing the data, training the models, model testing and validation. The climatological factors were used, as input values and Artificial Neural Network (ANN) implemented to establish prediction model. The model was developed through rapid application development (RAD) methodology. Upon completion it was deployed on a web environment to be used by various stakeholders in monitoring and predicting wild fires by giving a binary output of a yes or no on the likelihood of wildfire occurring. Artificial Neural Network Model was trained and validated using 80% and 20% of the set features respectively. The model gave performance accuracy of 82.69 per cent.