Прогнозирование пространственного поведения лесного пожара при неопределенности и нестационарности процесса

Abstract

The Russian forest fund, being a public domain of the people and a special kind of federal property, requires sustainable management at the national level. One of the key principles of forest management is to ensure that forests are conserved and protected against a wide range of threats, primarily forest fires. Although forest fires are a natural component of forest ecosystems and cannot be completely eliminated, researchers have currently revealed a decrease in the regulatory function and an increase in the destructive function of forest fires. Understanding the interrelations between the environmental factors and forest fire history is necessary for the development of effective and scientifically sound forest safety plans. The main purpose of the study is to increase the efficiency of the formation of an operational forecast of a forest fire in difficult conditions of a real fire (at instability and uncertainty). The author analyzed statistical data on forest fires the USA, Canada, Russia and the five southern European Union member states (Portugal, Spain, France, Italy and Greece) and confirmed the conclusion on the increase in the frequency of large forest fires. The most widely used in practice forecasting models of forest fire dynamics (Van Wagner, Rothermel, Finney, Cruz, etc.) and their computer implementations (Prometheus, FlamMap, FARSITE, VISUAL-SEVEIF, WILDFIRE ANALYST) are presented in the article. It is proposed to develop an intelligent system designed to create an operational forecast of a forest fire using convolutional neural networks (CNN). The structure of this system is described. It includes three main subsystems: information, intelligent and user interface. A key element of the intelligent subsystem is a forest fire propagation model, which recognizes data from sequential images, predicts the forest fire dynamics, and generates an image with a fire spread forecast. The scheme of the proposed model is described. It includes the following stages: data input; preprocessing of input data; recognition of objects using CNNs; forecasting the forest fire dynamics; output of operational forecast. The implementation features of the stage “recognition of objects using CNNs” are presented in detail: core size for each convolutional layer 3×3, activation function ReLu(x), filter in 2×2 pooling layers with step 2, max-pooling method, Object recognition and Semantic segmentation methods at the networks output.