Abstract
Multispectral archives contain evidence that RGB encoders cannot directly ingest. Sentinel-2 images include visible, red-edge, near-infrared, and short-wave infrared bands, whereas models such as CLIP expect three-channel natural images. We study image-to-image multispectral retrieval, where query and gallery images are encoded by the same method-specific pipeline and ranked by cosine similarity between descriptors. Starting from a generic RGB-pretrained CLIP backbone, with no Earth-observation pretraining, we propose a parameter-frozen test-time adaptation that updates only B band-weight logits per image. The method initializes the logits uniformly, refines them with a lightweight manifold-consistency objective steered by a class-name text prior, and reports band-level attribution scores. On 5-fold EuroSAT retrieval with disjoint query/gallery splits, all-band encoding raises Recall@1 from 80.28% for RGB-CLIP to 86.49% for our adapted descriptor and mAP to 46.50%; the all-bands-average reference is statistically indistinguishable on this benchmark, showing that the main EuroSAT gain comes from exposing the frozen encoder to every band. On a 5-seed BigEarthNet-MM evaluation, our method and the all-bands-average reference perform within one standard deviation across all metrics, while both surpass fixed spectral reductions and cache-based baselines by clear margins. We therefore frame the contribution as accuracy competitive with full-band averaging, but with per-band interpretability and near-zero adaptation overhead in the frozen generic-CLIP regime. Keywords: Multispectral retrieval, Vision-language models, Test-time adaptation, Spectral band weighting, Remote sensing