To address the challenges of low efficiency and poor robustness in multi-radiation-source term estimation within unknown environments, this paper proposes a variable state space-based particle filtering framework for multi-source term estimation. The framework dynamically constructs and updates the state space through the octree map integrated with radiation sensors. Efficient source term estimation is achieved via iterative observations and particle filtering algorithms. The global optimization capability of individual particles is enhanced through an adaptive differential evolution algorithm, while erroneous predictions are corrected by radiation distribution forecasting. Experimental validation involved Unmanned aerial vehicle (UAV) predictions and multi-algorithm comparisons across diverse scenarios, accompanied by systematic analysis of deviation sources. Results demonstrate that the proposed method effectively enables online inference of both radiation source parameters and their quantities from local multi-peak radiation fields.