Towards Audio-Visual Location Classification in Urban Soundscapes

Scene recordings captured in urban environments carry information not only about the type of environment in which they were recorded, such as an airport or park, but also about the specific city in which the recording was made. While scene classification has been extensively studied using audio, visual, and audio-visual cues, location classification, identifying the recording city, has to date been studied using only audio or only visual cues in isolation, leaving its audio-visual formulation unexplored. This gap matters because location-specific cues, such as region-specific acoustic characteristics and visual landmarks, are known to persist across different scene types within a city, suggesting that combining both modalities could meaningfully improve location recognition, and enable practical applications such as geotagging and location-aware urban monitoring. We address this gap by studying audio-visual location classification using two system formulations: a multimodal formulation, which combines audio and visual modalities, and a multi-task formulation, which learns location classification jointly with a related task (scene classification) through a shared backbone. For each formulation, we further examine early and late fusion strategies, differing in the stage at which audio and visual information is combined, and evaluate audio-only and vision-only unimodal variants to quantify each modality's standalone contribution to location classification. On the DCASE Audio-Visual Urban Scenes dataset, our best location classification result (67.63\%) is achieved by the multi-task formulation under early fusion, using OpenL3 audio and visual embeddings. Notably, we find that fusing audio and visual modalities does not consistently improve location classification over the strongest unimodal audio baseline (67.46\%, OpenL3 audio-only); most multimodal and multi-task configurations perform at or below this baseline, suggesting that combining modalities can dilute the strong location-discriminative signal already present in certain audio representations. As location classification using audio-visual cues has not previously been studied on this dataset, no established baseline exists for direct comparison; our results instead provide the first systematic benchmark for this task, and reveal that strong location classification performance depends more on the choice of a robust unimodal audio representation than on modality fusion itself.