Cloud-Robust Information-Seeking Agents for Adaptive Earth-Observation Acquisition and Surface Forecasting

Earth-environment monitoring increasingly has to deliver a decision inside a time window: an inundation map before a crest, a vegetation-condition field before an irrigation window, a damage map while a response team is still deployed. The observations that feed those products are neither free nor always available. Optical revisit is blocked by cloud, acquisitions compete for downlink and ground-station quotas, and synthetic-aperture radar (SAR) and in-situ stations carry different costs. Existing pipelines fix the acquisition policy in advance and spend their capacity on prediction, while recent language-model agents for Earth observation treat acquisition tools as cost-free black boxes and keep no auditable record of what each observation actually covers. We present CRISP, a cloud-robust information-seeking agent that decides what to acquire and then forecasts. A grounding ledger turns every observation into a typed entry with verifiable spatial, temporal, and quality fields and lets all downstream conclusions point back to it. A verbalized value-ofinformation critic converts the language model’s written reasons for a candidate acquisition into calibrated scores, which define a submodular cloud-gap coverage objective; a cost-normalized greedy rule then attains the usual (1 − 1/e) guarantee on a fixed budget and represents complementarity between optical and SAR acquisition. A physics-consistency self-critique recomputes advection, phenology, and radiation residuals, verbalizes the largest ones, and rewrites only the flagged time steps. On the EarthNet2021 surface-forecasting protocol CRISP reaches ENS 0.3075 on the out-of-distribution split and 0.2555 on the extreme split, improving the strongest official baseline by 7.7% and 8.1%; on SEN12MS-CR it reaches PSNR 25.62 and FID 86.2; on the flood-mapping and post-disaster protocols it reaches flood IoU 0.787 and mIoU 83.42. Under an equal acquisition budget it covers 0.78 of the cloud gap against 0.51 for an entropy criterion, and in an agentic protocol it raises task success from 62.4% to 78.6% while cutting invalid tool calls from 21.7% to 6.1%.