Using computer vision to monitor wildfire risk

According to a February report from the UN Environment Programme (UNEP) and GRID-Arendal, “climate change and land-use change are projected to make wildfires more frequent and intense, with a global increase of extreme fires of up to 14 per cent by 2030, 30 per cent by the end of 2050 and 50 per cent by the end of the century.” Certain regions will be more susceptible to these increases, including wildfire-prone regions in the western United States.

The same report found that 60 cents of every dollar spent in the U.S. on managing wildfires is put toward immediate firefighting responses: “Much less is spent on reducing fire risks in advance and helping communities recover in ways that could make them more resilient,” reports the New York Times.

Given this imperative, our data science team working with the US Department of Agriculture collaborated with the Accenture Federal Services Discovery Lab, a data science accelerator, to explore how federal agencies can use advanced technologies like computer vision (CV) to further mitigate wildfire risk and help protect our country’s natural resources and communities. We ultimately developed a CV model to identify dead and diseased trees in real-time using NASA’s Landsat, which can enable greater, more accurate visibility of tree mortality and therefore reduce wildfire risk. This model was recently recognized as a winner among 42 entries in the Amazon Web Services Disaster Response Hackathon.

The challenge

Identifying dead and diseased trees – large pockets of which can cause fires to start more easily, spread more rapidly, or even initiate fires by falling on power lines – is an essential tool to mitigating wildfire risk. Routine monitoring of the timing, location, and magnitude of tree mortality is critical for federal, state, and local land managers to rapidly respond to forest disturbances and understand the risk of future mortality.

Land management groups such as the Tree Mortality Task Force of California rely mainly on the Annual Insect and Disease Detection Survey performed by the United States Forest Service (USFS) to identify areas with tree mortality, according to external research. This survey provides a valuable asset for land managers, but our team saw an opportunity to augment that manual process with machine learning to expand the spatial extent and frequency of intelligence.

Creating a new model

We trained a CV model using 30m Landsat satellite imagery as our feature set. We labeled the level of tree mortality in each pixel of the Landsat imagery by converting shape-files from the Annual Insect & Disease Detection Survey Data into a Geo-tiff and transferring the labels to the Landsat imagery.

Utilizing a pretrained ImageNet, we created a feature representation of pixel change over time across multiple bands. The pixels were then classified as containing an area with or without tree mortality using a logistic regression with partial fit. Finally, we visualized our model output, error and ground truth using matplotlib and wrote the model output as a Geo-tiff.

This approach is extendable to any resolution or region of imagery. Ultimately, our model can assist land managers by providing consistent, spatially extensive measurements of tree mortality.

The model’s potential

We wanted to develop a model that allowed for near real-time assessment of tree health across the United States and provide a comprehensive assessment of tree mortality in a single snapshot. This model has several possible benefits:

1. Expanded visibility into population centers: The current technique used by the Forest Service covers the area of the Annual Insect and Disease Detection Survey. While survey regions contain many major forested regions within national forests, many forested areas on the wildland-urban interface are not included in the survey, and thus can be overlooked. In these more populated areas, increased wildfire risk due to tree mortality poses the greatest threat to human lives and livelihoods. Furthermore, cost of wildfire mitigation in these areas is higher. By providing spatially extensive measurements of tree mortality, our model equips land managers with greater insight to effectively reduce risk.
   
   
2. Reducing tree mortality through early detection of invasive insects that accelerate tree mortality: Early detection and continuous monitoring of established species such as the Western Pine Beetle are critical to slowing their spread. However, the low frequency and limited coverage area of the Insect and Disease Detection Survey reduces the capacity of USDA’s Animal Plant Health Inspection Service to rapidly detect tree mortality caused by outbreaks of invasive species. Due to the scalability of our CV model, federal, state, and local land managers can rapidly detect and monitor tree mortality over large land areas, leading to faster treatment of affected areas and containment of invasive species.
   
   
3. Reducing power line risk: Through faster identification and monitoring of tree mortality, our CV model has the potential to reduce the risk of power outages and forest fire outbreaks due to tree-conductor contact. On distribution systems, it is common for tree-related outages to comprise 20%-50% of all unplanned power outages. The vast majority of tree-related outages stem from tree failure; tree mortality exposes a power line to a high risk of tree incidents over time. By continuously identifying dead trees over a large land area, our model can enable utility foresters, asset managers, arboriculture consultants, and regulators to improve tree risk assessments and line clearance programs through automated detection of tree mortality.

Outlook for AI/ML applications for natural disasters

Our model ultimately shows how advanced technologies and the application of AI/ML can aid federal agencies in managing and mitigating complex issues by allowing greater visibility into and more efficient analysis of large quantities of data. AI/ML is a key enabler in transforming the vast amounts of planetary data into actionable insights that can be quickly delivered to decision-makers.

For this hackathon, we focused on building a model to aid in natural disaster mitigation, but we’ve also focused on community resiliency and emergency response as well as more efficient methods of geospatial analysis in the past.

Armed with this knowledge, federal leaders can make more informed decisions and more effectively allocate resources to carry out their multifaceted missions.

Thank you to the larger team who also contributed to this project: Cody Champion, Alex Farach, Alisha Kim, Jess Young, Ben Ortiz, Christian Conroy, Sam Kobrin, Merrall Echezarreta, Monica Puerto, Nima Latifi, and Ben Stone.