Using Sentinel-2 data to classify plants for better pollen forecasts
At Airmine we strive to make the best pollen sensors and pollen prediction models. Pollen levels are obviously dependent on local vegetation, more plants mean more local pollen spread.
Most people are allergic to specific types of pollen, such as grass or birch pollen, so we need to know where the allergenic species are located. To improve our pollen forecasts we are therefore classifying plants, starting with the Oslo area, Norway.
Based on images from Sentinel-2, we are identifying different plants and land use types, such as buildings and roads and areas covered by vegetation. The results are quite promising, we are able to pinpoint different tree types based on a modest amount of training data.
How it works
Very short: We use areas where we know what is growing combined with satellite images to train the machine learning algorithm. The algorithm can then be run on areas where we do not know what is growing. Lastly, we verify the results against another area where we know what is growing.
To train the algorithm, we have used a combination of public tree databases and manual mapping of trees.
We tested a couple of different algorithms, and ended up with the very fitting “random forest” algorithm.
Overall, we are very happy with the results and are expanding the model to cover other areas in Norway.
We get a fairly good accuracy for plant classification overall: 0.82. Our model is good at distinguishing between man-made structures (houses, roads) and vegetation, but struggles a bit more to to distinguish between the different conifers.
The next step is to use the plant classification in our pollen models, giving our customers more precise forecasts.
As a final note, we would like to send a big thanks to ESA Space Solutions for inviting us to speak at the COP26 Connectivity, Space and Digital Technologies for Green Value. We had the opportunity to present our air quality and pollen sensor and our latest work on plant classification.