Abstract: This paper proposes a method for forest fire spread direction analysis and ignition point backtracking to address forest fire investigation needs. The approach constructs a fire scene model through UAV mapping, utilizes a ResNet-18 convolutional neural network to identify fire spread directions from trace images, and generates a continuous direction field using inverse distance weighted interpolation. Runge-Kutta method is employed for backward streamline integration, with DBSCAN clustering ultimately determining the ignition point location. Simulation results demonstrate the method's effectiveness in analyzing spread patterns and tracing ignition points, overcoming limitations of traditional manual surveys such as low efficiency and subjectivity. By transforming expert knowledge into reusable intelligent algorithms, the approach reduces subjective dependence and achieves intelligent ignition point tracing, providing efficient, objective technical support for forest fire investigations.

Key words: forest fire investigation, spread direction identification, ignition point backtracking, convolutional neural network, backward streamline integration