QFlood: Low-qubit axis-aligned quantum attention for rapid flood mapping from multispectral and SAR observations

Flood mapping is a climate-relevant extreme-event task where satellite observations must be processed rapidly under limited labels and deployment constraints. We present QFlood, a low-qubit hybrid quantum-classical pilot for flood detection from multispectral and SAR Earth observations. Our earlier NASA Beyond the Algorithm Challenge prototype used a 6-qubit Quantum Vision Transformer and showed competitive flood-mapping performance with substantially fewer parameters than a SegFormer baseline. Building on this, we introduce QSKA-Hybrid, an axis-aligned quantum attention model that moves the quantum primitive from generic spatial patches to physically meaningful spectral and radar channel groups. QSKA-Hybrid uses a reusable four-qubit attention-coefficient circuit over band-group tokens while retaining a classical spatial branch for image-level capacity. On SEN12-FLOOD, a multimodal Sentinel-1/Sentinel-2 flood-detection benchmark, the model has 13,145 parameters compared with 13,613 for a matched ViT-Tiny baseline. In a three-seed pilot, QSKA-Hybrid shows directional F1 gains at 100 and 200 samples per class, with a reversal at 50 samples per class. We present this as a bounded low-qubit positive cell, not as a quantum-advantage claim. The main contribution is a hardware-aware evaluation pathway for weather and climate quantum pilots: align small quantum circuits with preserved physical Earth-data axes such as spectral groups, SAR polarizations, time steps, or subgrid variables.