Specialists from the Skoltech Center for Artificial Intelligence have created a machine learning-based system for predicting wildfires. Unlike its counterparts, the development analyzes dozens of parameters, from satellite data to population activity. The accuracy of forecasts reaches 87%, which allows preventing disasters before they occur.
The basis for training the neural network was archival data for 10 years, including information about the time and location of fires. The algorithm independently identified patterns, adapting to the characteristics of each region. Training the model for one subject took about a week. The system is currently being tested in Sakhalin, where it has already demonstrated its effectiveness in processing satellite data even in winter.
The key innovation is a comprehensive data analysis. AI takes into account the type of vegetation, vegetation index (reflecting the state of plants), relief, distance from roads and population density. Satellite images help track changes in the Earth's surface, and meteorological data helps predict risks five days in advance. For example, wind speed and air humidity can determine whether a random spark will turn into a large-scale fire.
Specialized indices play an important role, such as the Nesterov index, which calculates the balance of drying and moistening of the soil. However, the neural network complements traditional methods, using machine learning to improve accuracy. According to experts, access to detailed satellite images has become a breakthrough; previously, such data was not available.
Read materials on the topic:
На страже России: как искусственный интеллект помогает спасателям
Наводнения, пожары, объекты от 70 см – что и как видит российский «Ресурс-П» из космоса
В Уфе при помощи ИИ разработали систему предупреждения лесных пожаров