Building a model of early fire detection using machine learning methods and satellite imagery

Abstract

This work investigates the process of early fire detection by anomalous sources of thermal radiation registered by satellite remote sensing systems of the Earth. The subject of this study is anomalous sources of thermal radiation and their affiliation to an actual fire. The need for research is due to the fact that not every thermal anomaly recorded by a satellite corresponds to a real fire. Data on thermal anomalies were combined with the results of their ground verification in the Poltava oblast (Ukraine) in the period from 09/22/2020 to 04/23/2025. As a result of the combination, carried out on the basis of spatial and temporal proximity of points, 5096 clusters were established. At the same time, the share of clusters with confirmed fires was 13.0%. Each cluster was characterized by a set of 31 features, including spatial and temporal dimensions of the cluster, type of land cover, thermal radiation indicators, and meteorological indicators. A model has been built that determines the probability of a fire in a cluster for a given set of features. Using the method of recursive feature exclusion and a genetic algorithm, a set of 16 features was assembled, optimal according to the criterion of maximization of the PR AUC (Precision-Recall Area Under the Curve) metric.

A binary classifier has been built to identify fires among thermal anomalies. A feature of the model is its training on the results from ground verification of such anomalies. The model captures 85% of actual fires at a false detection rate of 54%.

In practice, this opens up opportunities for early detection of fire sites by ranking the detected sources of anomalous thermal radiation according to the criterion of fire probability. This makes it possible to improve the efficiency of using available forces and resources