Leveraging Satellite Onboard Computing for Real-Time Road Flood Monitoring and Dynamic Navigation

Core Concepts

Continuous monitoring of road flooding can be achieved through onboard processing of satellite imagery, enabling near real-time insights for dynamic navigation maps.

Abstract

The paper explores the feasibility of running road flooding detection models on-board satellites, using the simulated dataset and hardware constraints of the OrbitalAI Φsat-2 challenge.
Key highlights:

Developed a simulated dataset of multispectral satellite imagery over Bengaluru, India, annotated for flooded road segments.
Designed a ResUNet model architecture for water body segmentation, optimized for size and performance on edge computing hardware.
Tested the optimized model on a Jetson Tx2i board, achieving inference times of ~49ms per 256x256 image (1.48 sq.km. area).
Discussed the integration of the onboard flood detection with existing road networks to provide dynamic updates for navigation services.
Identified future work to extend the approach to a broader set of road health parameters and leverage multi-sensor data.
The proposed solution demonstrates the feasibility of leveraging satellite onboard computing to provide near real-time insights on road conditions, enabling more accurate and responsive navigation services.

Stats

The simulated dataset covers three distinct flood events in Bengaluru, India in 2022, with annotated water bodies ranging from 379 hectares before the flood to 776 hectares during the flood.
The optimized model size was reduced from the original 57.11 MB to 4.8 MB through pruning and quantization.
The average inference time on the Jetson Tx2i board was approximately 49.6 ms per 256x256 image (1.48 sq.km. area).

Quotes

"Continuous monitoring for road flooding could be achieved through onboard computing of satellite imagery to generate near real-time insights made available to generate dynamic information for maps used for navigation."
"With the capability of onboard computing performed on the latest satellite imagery, it is also expected that onboard processing will significantly reduce bandwidth costs of communication between earth observation satellites and ground-stations due to compressed information being downlinked."

Key Insights Distilled From

Continuous Monitoring for Road Flooding With Satellite Onboard Computing For Navigation for OrbitalAI Φsat-2 challenge

by Vishesh Vats... at arxiv.org 05-07-2024

https://arxiv.org/pdf/2405.02868.pdf

Deeper Inquiries

How can the proposed approach be extended to detect a broader range of road health parameters, such as damage, snow cover, or landslides, to provide a comprehensive view of road conditions?

The proposed approach of using onboard computing for satellite imagery to detect road flooding can be extended to encompass a broader range of road health parameters by incorporating additional machine learning models tailored to identify specific conditions. For instance, to detect road damage, a model can be trained to recognize patterns indicative of cracks, potholes, or structural issues on the road surface. Similarly, for snow cover detection, the model can be designed to differentiate between snow-covered and clear road surfaces based on spectral signatures unique to snow. Landslide detection can involve analyzing changes in terrain elevation and vegetation cover that may indicate potential landslide-prone areas.
To provide a comprehensive view of road conditions, a multi-task learning approach can be adopted, where a single model is trained to simultaneously detect multiple road health parameters. This approach can leverage shared features across different tasks, leading to more efficient model training and improved overall performance. By integrating various models for different road health parameters, the system can generate a holistic assessment of road conditions, enabling better decision-making for navigation and road maintenance purposes.

How could the use of multi-sensor data, such as radar or thermal imagery, enhance the accuracy and robustness of the road health monitoring system, especially in challenging weather conditions?

Integrating multi-sensor data, such as radar or thermal imagery, can significantly enhance the accuracy and robustness of the road health monitoring system, particularly in challenging weather conditions where optical imagery may be limited due to cloud cover or low visibility.

Radar Data: Radar sensors can penetrate through clouds and provide information on road surface conditions regardless of weather conditions. By combining radar data with optical imagery, the system can detect road conditions beneath cloud cover or during nighttime. Radar data can also offer insights into road moisture levels, surface roughness, and subsurface conditions, complementing the information obtained from optical sensors.

Thermal Imagery: Thermal sensors can detect temperature variations on the road surface, which can be indicative of snow cover, ice formation, or even landslides. By analyzing thermal imagery alongside optical data, the system can identify areas of potential danger due to icy road conditions or detect anomalies caused by landslides. Thermal data can also provide valuable information for predicting road surface conditions and optimizing maintenance efforts.

By fusing data from multiple sensors, the road health monitoring system can create a more comprehensive and reliable assessment of road conditions. The combination of optical, radar, and thermal data allows for a more robust analysis, improving the system's ability to detect and respond to various road health parameters accurately, even in adverse weather conditions.