Construction Method for Controllable Generative Dataset in Multispectral Methane Plume Segmentation

Multispectral methane plume segmentation is challenged by the scarcity of labeled samples, high data acquisition costs, and limited cross-scene generalization. To address these issues, this paper proposes a data construction and evaluation framework that integrates physical simulation priors, controllable generative expansion, and dual-scenario validation in both simulated and real-world environments. Based on Weather Research and Forecasting Large Eddy Simulation (WRF-LES) outputs, methane plume priors describing source locations, diffusion directions, and plume morphology are extracted to guide a physically constrained multispectral generation pipeline. By incorporating plume structure control and multispectral information embedding, the proposed framework enables the generation of diverse and physically consistent training samples. Its effectiveness is evaluated through quantitative experiments on simulated datasets and case studies of real methane leakage events. Results show that models trained with the controllable generated data consistently outperform those trained solely on WRF-LES simulation data across different segmentation networks, achieving FGIoU improvements of 19.9%–26.1%. Furthermore, the proposed method demonstrates superior performance in plume connectivity, diffusion-direction consistency, and background false-positive suppression, highlighting its potential for improving the robustness of multispectral methane plume segmentation under limited training data conditions.