Rapid disaster detection and response: A climate action collaboration

Integrating AI edge supercomputing with geo-comms for rapid disaster detection and response

By Soumik Sinharoy, Orange Silicon Valley; Shilpa Kolhatkar, NVIDIA; David Hammel, Balcony; Arthur Ecoffet, Orange; and Lorraine Chin, Orange Business Services

Editor’s note: This is a collaboration between NVIDIA, Orange Silicon Valley, and Balcony Labs. Our objective is to significantly improve the end-to-end process — from early detection of flood events, through precise and timely notification to the affected people, all the way to the effective orchestration of a dynamic response to reduce risk, accelerate response, and support the full response life cycle to maximize impact.

With unpredictable weather patterns caused by climate change, we are witnessing an increasing number of extreme weather events around the planet — from flash floods to severe droughts and wildfires in the Americas, Europe, Africa, Australia, and many other parts of the world. The ability to predict and respond effectively to disasters in near real time is a game-changer in preventing loss of life, property, and infrastructure.

In risk prevention and management, speed is of the essence to cross-check and confirm an event is occurring, to alert all potentially impacted parties, and to orchestrate the response. This coordination of information coming from various systems requires two major factors:

  • significant compute and data processing power;
  • the tools to make such data actionable in the hands of the decision makers, sometimes at random and unpredictable locations

Democratizing supercomputing capabilities is an important contribution toward enabling and reinforcing risk prevention. The combination of AI supercomputing at the edge of the network and targeted geo-intelligence-based bi-directional messaging with mobile phones at the affected sites is transformative. It enables the analysis of large amounts of data at the place it has been collected (at the edge), in areas that lack high-bandwidth connectivity. Such decentralized supercomputing makes the setup more resilient and efficient, and adheres to the principles of equality, data sovereignty, and privacy of different countries.

Earth observation satellites are the most effective instruments for the collection of geospatial intelligence data at a global scale, allowing for time-critical data gathering and post-processing for actionable insights. Specifically, satellites equipped with synthetic aperture radar (SAR) provide accurate earth observation capabilities — even when a storm is still ongoing — by monitoring weather-related hazards while penetrating clouds and other atmospheric conditions during day or night, allowing for the early identification of floods despite cloud coverage being prevalent over the affected areas.

Orange Silicon Valley collaborated with NVIDIA to test the capabilities of its AI on satellite SAR imagery. This was based on work done by citizen scientists Siddha Ganju and Sayak Paul on an AI inferencing model for flood detection. Orange Silicon Valley ran a performance benchmark on a system with dual-socket 20 Core Intel 5218 R processors with 10 NVIDIA A30 24GB Tensor Core GPUs and 768GB of system memory.

The model used for inferencing is an ensemble of three AI models, composed of two UNets and one UNet++, as outlined in a presentation at NVIDIA GTC 2021. As detailed above, SAR imagery was selected for this use case because it can be generated ‘through’ clouds without optimal atmospheric or light conditions, allowing near-real-time reaction.

The model throughput was evaluated using the NVIDIA Triton Inference Server. Two versions of the model were tested: a vanilla PyTorch model converted to TorchScript, and an optimized version using NVIDIA TensorRT. The batch size of 96 was chosen to maximize performance. TensorRT optimization yields up to 4x throughput improvement. With eight NVIDIA GPUs, it is possible to process up to 1,651 images per second. Given that conditional random field (CRF) post-processing can handle 1,326 images per second, using a harmonic mean calculation, we obtain a total throughput of 735 images per second.

Red indicates degree of flooding. Source: ESA

Each “tile” area of 63,000 square kilometers consists of 900 images, so one server at the edge of the telecom network with a direct (low-latency) feed of SAR images from a NOAA, NASA, or ESA satellite can scan a tile in 1.22 seconds and generate an alert of a potential flood event. The neural network model can be retrained to identify other natural disasters such as fires or tornadoes.

The outcome is a set of polygons representing the expected flooded areas as flooded or not. Severity is indicated by the gradient information.

Those polygons are then fed through an API into the Balcony geo-orchestration platform (detailed in the next paragraph) for immediate action with the goal of addressing the operational gap — where effective coordination at scale and speed before, during, and after a flood event has been an ongoing challenge for governments and local authorities. These challenges have resulted in misalignment of resources, delays in alerting and reaction time, and inefficiencies in reporting and data collection.

Balcony Labs has developed and fielded an integrated operations and critical event management solution based on its spatially-aware mobile-messaging technology, tied to mapping and spatial analysis. This system has been deployed to support operations in such nations as the US, Afghanistan, Mexico, Bangladesh, Haiti, and Ukraine. It enables scalable and real-time coordination of responders, field workforces, volunteers, supporting NGOs, and affected populations. Dynamic situational awareness is ensured by the integration of automation and AI tools. This system allows better and faster decision making with improved visibility, accountability, flexibility, and effectiveness for a broad spectrum of operations and emergency response situations.

Balcony’s geo-messaging addresses many of the challenges of real-time coordination through information exchange on dynamic locations. This enables flexible and rapid messaging flows linked to the place of physical activity, allowing one to quickly understand patterns and issues by deploying spatial analysis tools (that apply to satellite imagery for example). It connects multi-directional and interactive location-based messaging to real-time spatial data coming from remote sensor alerts. This creates a real-time digital space for engagement and management, which serves as a new kind of operating system for emergency response.

The mobile geo-messaging app is designed to populate ad-hoc any user base by using mass media, social media, and/or mass notification functionality to provide a link to download the app and join the location based ‘situational network’ for people in affected areas. Through the app, they can then report incidents, receive location-specific instructions, provide on demand insight to the decision makers, and maintain dynamic situational awareness as a function of their whereabouts. Unlike other messaging apps, geo-messaging does not require any contact information. It uses ‘privacy by design’ to ensure communications are location specific and not tied to anyone’s identity.

A final ingredient is necessary to combine all of the above into an end-to-end solution. Orange Business Services is capable of integrating all these pieces together, in many parts of the world, with a full-stack global connectivity solution, including the critical communications infrastructure. It can also draw insights at the edge to help rescue operations make timely and informed actions.

The complete workflow is as follows, where we bring together innovation from the previously mentioned collaborators to create a unique value proposition:

The example described above is very relevant in the context of climate change and the corresponding adaptation strategies. More generally, it is applicable to other contexts in both reactive situations such as wars and terrorist attacks, along with proactive initiatives such as informational support and training campaigns, which require the coordination of multiple teams and tools.

To learn more about Orange Business Service’s smart cities services and capabilities, please visit the Smart Cities page. Orange is working in collaboration with Balcony Labs and the NVIDIA Smart Cities program.