How are the forests in your country doing?
At Sogeti we love working with the latest technology and are passionate about saving forests. And a big portion of this is related to Data and AI. In this blog post our national Data & AI Lead, Marcus Norrgren, shows a concrete example where combining AI, tooling, scripting and traditional Geographical Information Systems (GIS) on a national level creates instant value.
There is so much we can do and solve with different solutions based on AI and the open world of data. With advanced image analysis, satellite imagery, Machine Learning, Deep Learning and AI methods we can for innstance hunt down the invasive spruce bark beetle and save our forests.
Let me give you a concrete example. We recently scanned all forests in Sweden. With AI-methods as a basis combined with smart tooling and other technical methods, we built a workflow, filtering, and aggregation of data to reach both micro-scale resolution and transformed the information into the "big picture" of things. And boom, created a snapshot of the current bark beetle damage across the country.
In the image, you see a map of Sweden divided into cells with a color coding. This coding is an index calculated to show which areas that our models indicate having been hit hardest by damages from the spruce bark beetle. This is based on actual damage predictions, not a risk map.
The heat is on
We have analyzed and summarized a whole country's forests for a specific damage signature to get a heatmap of damages. Sweden in this case, is just an example of, how we can create overlays at a selected area, even a country, and aggregate the data for authorities, companies and organisations, anywhere in the world, to help prioritize requiered actions due to their specific needs.
To achieve a bigger picture, it is important to not stare blindly on a small AI PoC project, but rather think about the problem you want to solve, how your output should be used wisely and make sure to puzzle the parts together. In this exmaple, we bring value looking at actual damage, not risk areas, and we can do it in a very short amount of time and for very large areas.
The how
We combine a unique pre-processing to achieve comparability of images over time, to get a high quality vitality map (showing where trees are dying). We combine this with a deep learning model that identifies forest species (spruce, pine, deciduous) and another one that is trained to detect spruce bark beetle damages. Thanks to a set of tools that we have developed, we can process and run models on images very efficiently. We can then filter our predictions to remove noise and irrelevant findings, as for example, remove all predictions in deciduous-dominated forests, since these will not spread or have a high risk of being false positives. We can also filter on tree height probabilities as well as the amount of observed tree vitality change.
The output, based on the process described, will result in points or polygons with probabilities and some other metadata, showing where our models predict spruce bark beetle damages and how accurate the predictions are.
We do acknowledge the challenge with validation, and we also acknowledge, as for any model, predictions are never perfect. However, the heatmap lines-up quite neatly with areas we know have problems. The northern ones need more investigation, but we have done field checks and found spruce bark beetle attacks also far north, which is a bit worrying.
However, if you have many predictions, or large areas and need to prioritize how to act, specific points will not help. You have to get an aggregated view. For this, we created a script which can summarize geo information into a given shape (stands, grid, etc.) and then summarize the number of damages and size of damages into an index (as shown in the image below).
The index shows the grade of damages based on size of damages in square meters and the number of damages. Indexes can be defined in many ways, and the above example shows more of an abritrary number for the heatmap and not an actual unit which of course is possible.
So, in short, we can find unique signatures using remote sensing imagery. We can filter it using other models and methods. We can then summarize it into a high-level picture for prioritization, and then provide details for a given area.
Sounds interesting? I am already drafting my next blogpost on the topic. Cant't wait? Reach out to me to discuss your challenges and possibilities.
- Marcus NorrgrenData & AI Lead, Sverige
073-061 96 80
Marcus NorrgrenData & AI Lead, Sverige
073-061 96 80