A Test of Habitat Selection at Two Spatial Scales in a Sitand-Wait Predator: A Web Spider in the Namib Desert Dunes

Aerial detection requires fine detail for small targets and broad context for large structures. Sampling loss, blur, and noise can weaken detail without changing feature-map resolution, creating a demand-reliability mismatch. We present Physical Imaging Reliability-Guided Scale-Biased Feature Reweighting (PIR-SBFR), a sensor-to-inference interface that separates scene-scale demand from observation reliability. A visual branch estimates demand, while a fixed analytic rule converts relative sampling, sharpness, noise, and field availability into an ordered multiscale prior. The routing rule requires no sensor-specific response fitting. On public DIOR images, ten paired runs increase average precision (AP) from 62.98% (SD 0.32) to 65.52% (SD 0.40). On the study-specific AI-TOD-v2 split, AP rises from 29.58% (SD 0.27) to 32.10% (SD 0.41). Matched local reproductions of three contemporary tiny-object detectors reach 30.42-31.59% AP under the same protocol. PIR-SBFR also improves three compatible host detectors by 2.09-2.54 percentage points. All 27 controlled quality conditions and nine held-out degradation types favor PIR-SBFR. On a Jetson Orin NX, its FP16 end-to-end median latency is 19.55 ms. These results demonstrate effective acquisition-conditioned routing under controlled observation changes on public aerial imagery.