Satellite On-Orbit Chip-Level Deep Learning Model for Real-Time Dust Storm Monitoring.

Dust storms evolve on 10 min time scales, yet satellite monitoring remains constrained by ground segment pipelines that delay actionable products by half-hour to hours. We present an on-orbit deep learning framework that executes dust storm detection and quantitative retrieval directly onboard the satellite, converting geostationary Himawari-8/9 observations into exposure-grade products within 5.62 min─an ∼80% latency reduction relative to conventional acquisition-to-product chains. A cascaded design couples a lightweight event gate with a multitask retriever for PM10 and PM2.5, optimized by a tail-aware loss that prioritizes fidelity at extreme concentrations. Compared to the baseline among LGBM, XGBoost, MLP, CNN, and LSTM, our system reduces the RMSE by 30% (from ∼32.25 to 20.54 μg/m3) for PM10 and by 25% (from ∼14.44 to 10.27 μg/m3) for PM2.5, markedly mitigating the underestimation of high-concentration values. To assess on-orbit feasibility, we simulate deployment on an NVIDIA Jetson AGX Orin platform, representative of resource-constrained satellite computing environments, where the model achieves a 3.60 s inference latency with a ∼10 W power draw and <3 GB memory footprint, confirming its viability for resource-constrained spacecraft environments. Case studies (May 19-20, 2023; April 15, 2025) show spatially coherent dust cores, sharper plume gradients, and minutes-scale fusion with population grids and vulnerable locations for real-time exposure assessment and early warning. This work integrates on-orbit inference, tailored retrieval, and demographic coupling to transform dust storm monitoring into minute-level, exposure-focused intelligence, providing a scalable model for other disaster aerosols.