Spatiotemporal Conditional Diffusion with a Vision-Language Model: Spatial Quality Modeling for Adaptive Fusion of Radar Forecasts

Combining diffusion forecasts requires decisions about which candidates to retain and how to preserve strong echoes. We propose STCD-VLM, which learns candidate subset quality for adaptive fusion of radar nowcasts. Six observations condition spatiotemporal conditional diffusion (STCD) to generate five candidate sequences covering 120 min. Radar descriptors and representations from a frozen Gemma 4 31B vision-language model (VLM) condition a shared quality network. The network evaluates all 31 nonempty subsets and selects one for the complete sequence, combining its mean and pixelwise maximum. Evaluation on 256 event groups from eight Chinese S-band radar collections yields critical success indices of 0.25871 and 0.12712 at 30 and 40 dBZ. Relative to five-candidate maximum aggregation, these scores increase by 0.04108 and 0.02185, while false alarm ratios decrease by 0.08070 and 0.08203; 15 km fractions skill scores improve by 0.08433 and 0.07822. Fixed five-candidate fusion accounts for most of the mean CSI improvement over maximum aggregation. Adaptive selection adds 0.00170 to this mean under the same fusion rule, with higher CSI30, lower false alarm ratios and slightly lower CSI40. A matched zero-VLM control has similar principal score estimates, leaving the independent benefit of VLM inputs to be established. These comparisons support a quality-guided fusion framework that improves strong-echo overlap and neighborhood coverage agreement relative to maximum aggregation, while producing an event-specific forecast through one sequence-wide subset decision.