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

TL;DR

This paper explores the feasibility of onboard satellite image processing for real-time road flooding detection, demonstrating that low-size, high-accuracy models can be developed for use in navigation systems.

Contribution

It presents a novel approach to onboard satellite computing for flood detection, including dataset creation, model development, and validation for the PhiSat-2 mission.

Findings

High-accuracy flood detection models are feasible onboard satellites.

Simulated dataset and annotation process for Bengaluru floods are established.

Low-size models suitable for onboard deployment are developed.

Abstract

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. Given the existing computing hardware like the one considered for the PhiSat-2 mission, the paper describes the feasibility of running the road flooding detection. The simulated onboard imagery dataset development and its annotation process for the OrbitalAI {\Phi}sat-2 challenge is described. The flooding events in the city of Bengaluru, India were considered for this challenge. This is followed by the model architecture selection, training, optimization and accuracy results for the model. The results indicate that it is possible to build low size, high accuracy models for the road flooding use case.