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

Area of Science:

• Earth and Space Science
• Artificial Intelligence
• Environmental Monitoring

Background:

• Satellite monitoring of dust storms faces significant latency issues due to ground segment processing.
• Existing methods delay actionable products, hindering timely public health and safety responses.
• Rapid dust storm evolution necessitates near real-time monitoring capabilities.

Purpose of the Study:

• To develop and validate an on-orbit deep learning framework for real-time dust storm detection and quantitative retrieval.
• To significantly reduce the latency of dust storm product generation from hours to minutes.
• To provide exposure-grade dust storm data for enhanced public health and disaster management.

Main Methods:

• Implemented a cascaded deep learning architecture with an event gate and multitask retriever for PM10 and PM2.5.
• Optimized the model using a tail-aware loss function to improve accuracy at extreme dust concentrations.
• Simulated on-orbit deployment on an NVIDIA Jetson AGX Orin platform to assess feasibility in resource-constrained environments.

Main Results:

• Achieved an ~80% latency reduction, delivering products within 5.62 minutes.
• Reduced RMSE by 30% for PM10 and 25% for PM2.5 compared to baseline models, significantly improving high-concentration estimates.
• Demonstrated on-orbit viability with low power draw (∼10 W) and memory footprint (<3 GB).

Conclusions:

• The on-orbit deep learning framework offers a scalable solution for minute-level dust storm monitoring and exposure assessment.
• This approach transforms satellite data into actionable intelligence for real-time early warning systems.
• The methodology is adaptable for monitoring other atmospheric disaster aerosols.