Community Perception of Wildfires: An Evaluation of Algorithms for Detecting Visual Elements from a Territorial Dataset

The wildfire management involves the required activities to reduce both the risk and the impact, but the territorial knowledge and experience must be considered. An image dataset was developed by users contributions through a mobile application. Nevertheless, it is required to analyze the common visual characteristics against geographical nearest groups to improve the territorial decisions. The aim of this work is to evaluate the accuracy of a set of algorithms based on image learning techniques, by considering the model architecture and, the label selection according to the wildfire domain. We assess the accuracy of transferring learning from pre-trained models in computer vision tasks to a particular domain. We apply transfer learning techniques to leverage prior knowledge and representations learned by the pre-trained models. The algorithms are developed and evaluated using a domain specific dataset developed by a community. Such dataset is highly related to forest fire images available on the citizen science platform E-ncendio, which is part of the FONDEF ID22I10072 project. Classic Convolutional Neural Network models have been implemented, such as Inception-V3, Densenet, VGG-16, VGG-19, and the Residual Neural Network ResNet, which have proven to be capable of identifying visual elements that represent objects present in the landscape subject to forest fires. This allows the algorithms to be adapted to characterize community consensus, providing a territorial understanding of the perception of forest fire risk, as well as factors related to its prevention, response, mitigation, and recovery. This automated approach has the potential to be applied in other fields that require the identification of visual elements in images, broadening its impact in various areas of research and development.