Continuous Monitoring for Road Flooding With Satellite Onboard Computing For Navigation for OrbitalAI {Phi}sat-2 challenge

Overview

• The paper explores the feasibility of using onboard computing of satellite imagery to detect road flooding in near real-time and provide dynamic information for navigation maps.
• It describes the development of a simulated onboard imagery dataset and its annotation process for the OrbitalAI {Phi}sat-2 challenge, focusing on flooding events in Bengaluru, India.
• The paper presents the model architecture selection, training, optimization, and accuracy results for this road flooding detection use case.
• The key finding is that it is possible to build low-size, high-accuracy models for real-time monitoring of road flooding using satellite data.

Plain English Explanation

The paper looks at a way to use satellite images to quickly detect when roads are flooded. This could be done by having computers on the satellites analyze the images as they are taken, rather than sending the images back to Earth first. This could provide near real-time updates on flooding for navigation apps and maps.

The researchers created a simulated dataset of satellite images showing flooding in Bengaluru, India. They then used this dataset to train and optimize a machine learning model to detect road floods from the satellite images. The process of developing and annotating this dataset is similar to the approaches used in the SpaceNet 8 challenge.

The results show that it is possible to build small, accurate models that can run on the computers inside satellites to detect road flooding. This could allow for real-time flood monitoring and updates to navigation systems, without needing to send all the satellite images back to Earth first. The ability to process data on the satellites themselves, rather than centrally, is an active area of research in remote sensing and edge computing.

Technical Explanation

The paper explores the feasibility of using onboard computing of satellite imagery to detect road flooding in near real-time. It describes the development of a simulated onboard imagery dataset and its annotation process for the OrbitalAI {Phi}sat-2 challenge, focusing on flooding events in Bengaluru, India.

The researchers created a dataset of simulated satellite imagery showing road flooding in Bengaluru. They then annotated this dataset to label the flooded areas. This process of creating and annotating remote sensing data for machine learning models is an active area of research.

The paper then presents the model architecture selection, training, and optimization process for this road flooding detection use case. The results indicate that it is possible to build low-size, high-accuracy models that can run on the computing hardware available on satellites, like that used for the PhiSat-2 mission. This aligns with broader research on processing sensor data at the edge, rather than in a centralized location.

Critical Analysis

The paper provides a promising proof-of-concept for using onboard satellite computing to enable near real-time road flood monitoring. However, the research is limited to a simulated dataset and a specific geographic region. Further testing and validation would be needed to ensure the approach works reliably in diverse real-world conditions.

Additionally, integrating this type of flood detection system with dynamic mapping and navigation applications would require additional research and development. Challenges around data latency, coverage, and integration with existing systems would need to be addressed.

While the paper demonstrates the technical feasibility, more work is needed to understand the operational, business, and policy implications of deploying such a system at scale. Automated mapping of urban features from satellite imagery is an active area of research with wider applications.

Conclusion

This paper presents an innovative approach to leveraging onboard satellite computing for near real-time road flood monitoring. By performing the image analysis directly on the satellite, rather than transmitting all the data back to Earth, the system could provide dynamic flood information to navigation apps and maps much faster.

The researchers have shown the feasibility of building small, accurate machine learning models capable of running on existing satellite hardware. This is an important step towards enabling satellite-based flood monitoring systems that can deliver timely, actionable insights to support disaster response and transportation planning.