On behalf of the Department of Public Works, Aquila Ecology investigated the possibility of surveying road verges from a moving car. With this project, we are testing the limits of what is possible with taking images and analysing them with AI. And with results. Although there are still plenty of challenges in recognising plants and vegetation types, it already appears to be quite possible to recognise larger plant species. For instance, we were able to detect 4 different exotics on a short trajectory: Japanese knotweed, Sky tree, Giant hogweed and Pampas grass, while the two trained ecologists in the car noted only Japanese knotweed and Giant hogweed. We do this by taking pictures with a short shutter speed while driving at 80 - 100 km/h. These photos are then analysed by AI.
Photographing roadsides
The photos that can be taken from a moving car using this method are of surprisingly high quality. Even small species like Soft storkbill can be clearly identified in the photos. To achieve this, we tested two different methods of shooting and also created, trained and tested a GAN (Generative Adversarial Network) to counteract blur. This technique is familiar from websites that allow whole images to be composed based on some text, but can also be used for image enhancement. We also built a platform that can be slid out and in by the driver of the car, with the camera mounted on it, which can also be pointed up and down. Thus, in principle, one person could survey roadsides.
Recognition with AI
We took about 150 photos per kilometre, which were then analysed by AI. This showed that large species were recognised well, but smaller species were not always identified correctly. Very likely, the training data behind the AI model does not sufficiently match the image taken from the car. A quick test already showed that recognition could be greatly improved by improving training data.
Although far from all plants are recognised correctly, the amount of data is so large that general patterns can be extracted from them: the correctly recognised plant species are often represented enough to compensate for the incorrect recognitions. Using a method developed by Aquila Ecology, we converted the recognised plant species to vegetation types according to the RWS typology. This showed that the correct vegetation types could be predicted with around 50% accuracy. The data can be read directly into GIS programmes.
Application
With the developed technique, it is possible to monitor hundreds of kilometres of roadside at low cost. At least in the process, the distribution of large exotics and other large, recognisable plants can be properly imaged. In time, it will probably also be possible to obtain an even more detailed picture of the vegetation, down to the presence of grass species such as glossy oats and whitebell.
