AeroReformer2: Spoken-query referring segmentation for aerial images

Spoken queries provide a natural, hands-free way to identify targets in dense remote-sensing imagery. Existing referring segmentation benchmarks in this domain accept only written expressions. We introduce VoiceAeroRef, a spoken-query benchmark derived from RISBench. It adds synthetic speech with diverse accents and voices while preserving the source images, masks, and data splits. A hard test set combines rotor, wind, and mixed interference with three signal-to-noise levels. We also propose AeroReformer2, a bilateral network for efficient speech-conditioned segmentation. The network retains speech tokens and conditions visual features at two scales through kernel linear attention. This avoids constructing a dense speech–visual affinity matrix. A detail path and a high-resolution refinement head preserve object boundaries. On the clean test split, the Swin-Base version achieves a mean intersection over union (mIoU) of 62.09% and an overall intersection over union (oIoU) of 68.22%. These scores exceed those of the strongest audio-adapted remote-sensing baseline by 5.38 and 2.08 percentage points, respectively. The model also achieves the highest mIoU on the hard test set, at 54.09%. To the best of our knowledge, this is the first study to introduce a benchmark and model for full-sentence spoken-query referring segmentation in remote-sensing imagery. The project, along with code and dataset, is available at https://lironui.github.io/AeroReformer2/ .