Monday, April 22, 2024

Machines help learn Gabrielle’s full impact

Avatar photo
Scion researchers have retrained their AI mapper to paint a picture of the cyclone’s effects on forestry, and forestry’s contribution to flood damage.
BEFORE AND AFTER: AI-generated forest boundaries pre and post-cyclone on a tributary of the Tutaekuri River, Waiwhare, Hawke’s Bay. Illustration by Melanie Palmer (Scion, Remote Sensing & GIS) using imagery sourced from the LINZ Data Service and licensed by Hawke’s Bay Regional Council for reuse under CC by 4.0.
Reading Time: 3 minutes

Some smart tech work by Scion researchers has taken AI modelling software developed to count new tree plantings, and tipped it on its head to instead determine tree losses from Cyclone Gabrielle.

As the forestry sector wrestles with its right to operate in catchments severely damaged by forest waste and tree losses into waterways, the work is providing a means to also fully understand and measure forestry’s full impact.

Scion remote sensing scientist Grant Pearse of the agency’s Digital Forest Project team said they were prompted to pivot from mapping tree plantings to mapping Gabrielle’s effects in Gisborne and Hawke’s Bay as forestry damage became clearer. 

It is recognised as the first opportunity to use digital artificial intelligence tech to determine the impact of a natural disaster on New Zealand’s planted forest estate. 

The tech combines aerial images with AI and deep machine learning programmes that can accurately detect, and map planted forests using RGB (red, green, and blue) imagery technology. 

RGB captures natural colours that are visible to the human eye, so provide easily interpreted information within the images it captures.

At present the model targets pine and Douglas fir and can map stands two to three years post-planting, ranging in scale from commercial forestry to small woodlots and even farm shelter belts.

Pearse says the granularity of the technology can be very fine, down to a few trees or even a single tree if resolution is high enough, but generally forest areas are mapped to around 1500sqm.

Tairāwhiti was hit hardest by cyclones Hale and then Gabrielle in short succession, with rapid assessment showing many closed-canopy forests suffered significant damage and landslides. Significant amounts of debris remain on slopes or on river margins in the region.

Working with Gisborne District Council, the researchers can now assess the loss of net stocked area of the region’s forests, given that the normal stock area in a forest is already known. 

The resulting difference provides a good indication of the wood debris likely to migrate downstream over time. Establishing the volume enables councils to identify and quantify the risk in specific catchments and plan ahead to mitigate that risk.

The team has just added Hawke’s Bay forests to the dataset. The work has provided insights into where forest was standing prior to Gabrielle, what species were in place and what has been lost.

A recent Landcare report highlighted the vulnerability of exotic forest cover as a contributor to landslide risk, particularly through the Gisborne region. 

That work found exotic forests’ ability to reduce landslide risk deteriorated the further north one went from Wairarapa through to Tairāwhiti. In Tairāwhiti exotics were estimated to be only 50% effective at reducing landslide risk, compared to a 90% reduction in risk under indigenous cover.

While repurposed for the cyclone work, the AI model’s original intent remains, namely to develop a digital template of the country’s exotic forest inventory.

Further ahead, the Scion researchers aim to combine their AI tech with LiDAR (light detection and ranging) sensing that uses pulsed laser light to measure distances to the earth. 

This will enable them to better determine tree attributes including age class, stand density, timber volume and carbon accumulation. Gisborne will be the first region to trial this tech package.

Partnering with Land Information NZ and its existing LiDAR data warehouse, the country will soon have a national scale “digital twin” of its productive forest estate. 

From there remote sensing will enable monitoring and mapping of forest health, productivity and carbon capture. Modelling of possible future forest scenarios will also be made possible.

Other possible benefits of the tech are combining it with data from other agencies that collect the likes of rainfall data and have soil mapping information that could provide an even broader landscape-based insight to how the land is influencing forests’ behaviour under extreme conditions.

People are also reading