In buoyant plume environments, the highly unsteady nature of gas diffusion direction invalidates the conventional assumption of steady-state diffusion direction underlying information-driven algorithms. This leads to inaccuracies in information entropy gradient estimation, causing erroneous particle convergence. To address these issues, this study proposes an improved infotaxis source-tracing method incorporating exponential weight adjustment and stratified resampling. First, an exponential weight correction coefficient is introduced during weight normalization to effectively balance weight distribution disparities and reduce erroneous convergence of particles. Then, a stratified resampling model based on residual information is established, which employs quantile-based stratified sampling to improve source localization accuracy. Finally, a dual-threshold detection mechanism using proximity index-block sequences is developed to trigger position migration upon detecting search stagnation, thereby effectively overcoming local optima. Simulation results demonstrate that the proposed strategies significantly enhance the algorithm's robustness and search efficiency in unsteady diffusion environments.