Big Data: What took many man-hours of laborious and costly sampling work in the past and yet was barely enough to approximate the true condition of forests, will be collected as a complete record by (partially) autonomous technology over the medium term and threatens to overwhelm us like an avalanche of data. This was the impression one could get at the Austrian Federal Forests’ Research Day, which took place on October 2 in Salzburg. Other thematic highlights were technologies for recognizing the physiological properties of forest plants, safety training with extended reality methods as well as future areas of application for robots and autonomous wood harvesting machines in forests.
East Tyrol shows how it’s done
In order to speed up the creation of a forest cover on clearings caused by various calamities, East Tyrolean forest managers used drones in difficult or inaccessible terrain as a way of sowing seeds of pioneer tree species rowan and birch as well as stand-forming tree species larch and pine. In order to protect the seeds, increase the germination rate and adjust the weight for better distribution, the Dronenring team used pelleting – a type of coating of seed packets with clay minerals. A total of five areas in East Tyrol were selected several months earlier. The drone was flown when there was a suitable snow cover in order to be able to check how evenly the seeds were spread, as Drohnenring’s Stefan Hölzl-Strohmayr explained. The selected cargo drone with a payload of 30 kg, which has been used in agriculture for years, was able to spread seeds on up to 50 hectares during a 25-minute flight, according to the drone expert. Studies on germination rates on different types of soils and exposures are currently underway.
Shortcut to drought resistance
BFW’s Marcela van Loo talked about the possibility of selection in seeds
Dr. Marcela van Loo presented the method of phenotyping, which can be used to identify the desired drought-resistant origins of a tree species even in small forest plants. The precise appearance of plants is determined with the help of 3D, RGB, thermal imaging, hyperspectral and chlorophyll fluorescence cameras in a climate-controlled environment in a room-filling facility at the Vienna BioCenter. Each seedling is weighed during each measurement run and is given a different amount of water depending on its weight and assigned climate scenario. As the scientist explained, the climate chamber then shows how the seedling copes with the respective climatic conditions.
Robotic dogs in the forest
© Robert Spannlang
As part of the DigiForest project, an international team of researchers is exploring ways of creating high-resolution data at tree level by means of robots which move autonomously on a set route. After the digital identification by foresters based on the collected data, unmanned wood harvesting machines are supposed to be able to automatically find and fell the identified trees. Lidar laser scanners are used here, which can also be mounted on drones or carried through the forest in backpacks or on harvesting machines, as robotics expert Stefan Leutenegger from the Technical University of Munich (TUM) explained.
In order for drones to be able to avoid trees in the forest even without GPS, AI, i.e. an artificial neural network, specially trained for this application is used. Autonomous drones without expensive lidar technology have also been built at the TUM. They create infrared and depth maps with cameras which help them avoid even fine branches and fly independently along a predetermined route through the forest. “Some of the main goals for the future are to let autonomous drones fly through the forest even faster in order to increase efficiency, as well as the reliable extraction of parameters such as the tree species,” the TUM scientist emphasized.
Automated expert knowledge
“It takes sawmill employees just a few seconds to assess a tree that has grown for an average of 120 years,” ÖBf logistics manager Wolfgang Holzer said, highlighting the status quo of the quality assessment of log wood. In order to make this process somewhat more objective and understandable, an extensive database for the development of an AI-based system is to be created as part of the MeRu project. For this purpose, 2,000 logs were scanned at a pilot plant at the ÖBf’s log yard in Amstetten. The logs’ features, such as cracks and discoloration, were assessed by experts from the forestry and sawmill industries, and their assessments were fed into the AI database.
“The data pool for training an AI is available and stable. We are also trying to establish a kind of roadmap towards possible standardization,” emphasized Holzer. Project partners include well-known companies such as Microtec, Felix Tools, the Austrian Institute of Technology, and the Holztechnikum Kuchl.
Forest inventory – digital and automated
“The traditional forest inventory is a very labor- and cost-intensive business,” Arne Nothdurft, Professor for Forest Monitoring at the University of Natural Resources and Life Sciences (BOKU) in Vienna, said at the beginning of his presentation. In addition, the data flow is interrupted several times and reproducibility is rather low. “In contrast, laser scanning systems have 10,000 to one million measuring points per second in 3D space,” the Chair of the Institute of Forest Growth explained. It is rather challenging to use this data cloud to segment individual trees using cluster analyses based on the locations with clusters of points. On a test circle, each tree can be reassembled from the points assigned to it and then measured according to all kinds of criteria: tree height, trunk diameter at any height, crown volume, crown projection area and even the calculation of the biomass which can be used for energy generation. Even tree species recognition can be done with a very high degree of accuracy in this process.
Another application of laser scanning is the fast and systematic estimation of the volume of damaged wood after storms. “When storm Vaia swept across the Gail and Lesach valleys in 2018, we came up with the idea of bringing this digital forest inventory to the area together with the BOKU Institute of Forest Engineering. So, we started to estimate the windthrown trees drawing on laser-based inventory data,” Nothdurft said, adding that this does not even require an aerial survey. Based on measurements on those parts of the stand that are still standing, regression and stochastic models can be used to draw conclusions about the stock that is no longer there. The resulting maps showed how much damaged wood accumulated in the affected areas with associated error probabilities.
Based on terrestrially collected laser data, relatively accurate spatial estimates of a forest’s structure can also be derived – for example, arbitrarily defined diameter categories in protective forests. These, in turn, indicate whether there is a need for action with regard to the distribution of chest-height diameters of trees in order to estimate the stability of a protective forest, as the forest inventory expert emphasized.
Aerial forest monitoring
In contrast to aerial photography and lidar systems mounted on aircraft, drone-based, three-dimensional laser scanning (Lidar) has a very high point density, thanks to which structures underneath the tree crowns – such as regeneration – can be recorded as well, Phillipp Fanta-Jende from the Austrian Institute of Technology (AIT) explained. This is also made possible by the much lower flight altitude of the drone, which enables the drone to do diagonal scans in front of and behind it.
“The aim is the automatic classification of tree, shrub and herb layers. In order to train an AI to do this, you would need a massive amount of data annotated by forestry experts. This is why synthetic training data are also included,” the AIT researcher said. Typical basic patterns of tree species were extrapolated over the rotation period using data from the yield tables, randomized according to certain criteria and expanded to include simulated lidar scans. The question now is: How many synthetic trees can a model tolerate in order to achieve sufficient accuracy? “Ultimately, it’s a question of economic efficiency,” Fanta-Jende admitted and added: “This approach is sufficient to train an artificial intelligence, but it would be a botch for real-world validation.” There are already promising model results which allow the easy derivation of stratification, the scientist concluded, ending with the sentence: “We are currently working on a hydrogen drone that takes off in Vienna, lands in Salzburg and scans everything in between.”
