Harnessing AI for forest conservation and berry picking 

Artificial Intelligence is making an entrance to our everyday life. In scientific research we are looking for new ways to utilize AI for the benefit of society and academia.  

This is a challenging task, considering that the essence of research is to be curious and try new things boldly – but also ensure reliability, safety and security of any data.  

At the FGI, we have already built a map that predicts the presence of berries in certain areas. The prediction is using certain environmental parameters, such as vegetation/forest type and elevation of the area. This July, our scientists travelled to the area where the pilot version of the map has been drawn, to get in-situ field data of the berry yield. 

Each study site had 10 test plots that were photographed with a camera and a drone. The berries of the test plots were calculated, to get validation data for the automatic detection algorithms. The work is carried out in the FEROX project, funded by the European Union 

Deep learning methods added to the currently existing berry maps 

The data scientists tested berry recognition and detection algorithms on the photographs, taken on-site at the FGI berry map area, as well as the drone-survey images. They compared the performance of various deep learning algorithms on both types of images using two different schemes of training the deep learning models:  

1. fully supervised training 

2. semi-supervised training.  

The researchers found that estimating the berry yield from on-site photographs and drone-survey images can be done with high accuracy, even in real-time. It is made possible using consumer grade hardware for both the data collection and computation.  

The results were unexpectedly positive. Especially given the relatively small size of the berries in the images compared to more typical objects in computer vision problems, such as vehicles. Such objects are generally much larger than berries, which makes their detection easier. 

Deep learning models are traditionally trained in a fully supervised manner. The training examples (such as the locations of individual berries in the images) are explicitly given to the training process by a professional annotator. The annotator will have domain-specific knowledge of the problem.  

Unfortunately, manually annotating large amounts of berry data in this way is inconvenient, due to the high numbers of individual berries found in the images. Therefore, the researchers also tested a semi-supervised training approach:  

• automatically locating interesting objects from the images, 

• only then explicitly telling the deep learning model was explicitly whether berries were the found objects.  

The proposed training method did not require the researchers to annotate the locations of each individual berry in the training images. This was done automatically by the model, substantially reducing the burden of data annotation. 

The key task is in making the data annotation more efficient   

The goal is to train our berry prediction model to perform even better in highly diverse forest and bog conditions. Additionally, we are researching multiple different ways of training our models with less human annotated data.  

Obtaining large amounts of high-quality training data for deep learning models has been problematic. This is because there are only a few professional data annotation companies globally that have the required knowledge for distinguishing between different berry species and their ripeness levels.  

We are hoping to further improve our semi-supervised data annotation process, where deep learning does part of the work. From this, we hope that the end-user would not need to manually annotate thousands and thousands of berries from the survey images.  

The methods developed for automatic berry recognition and detection could also be used in numerous other forest related applications. These methods could apply where real-time computer vision algorithms are needed. However, only a few established pre-annotated large-scale datasets – necessary for training the models – currently exist.  

Harnessing artificial intelligence to manage the vast amounts of other forest data   

Meanwhile, FGI also has scientists working on similar detection algorithms for the whole forest. The UNITE competence centre has created technologies that allow building an accurate digital replication of boreal forests at individual tree level, using aerial laser scanning surveys of Finland as source.  

The technology is demonstrated in the Metsäkanta forest database, which enables users to view individual trees sorted by length and tree species. The system is open to test users. The results have also contributed to automating field sampling of forests to be several dozen times faster than earlier. 

The algorithm that identifies wood species is now in operation, and it is estimated that the area laser scanned by the National Land Survey in Southern Finland in 2020–2022 will be calculated by the end of this year. There are an estimated 2 billion trees in this area, the diameter, species and volume of which will soon be viewed from the forest database. You can read more about the work carried out at National Land Survey of Finland with the digital twins in a separate article. 

This detailed data on the individual trees, species distribution and trunk size can be helpful in determining areas that are valuable for forest conservation purposes. Laser scanning surveys combined with data collected, e.g., by sensors attached to forestry machines, produce impressive amounts of accurate, up-to-date data on our forests. Currently, the bottleneck in research is more present in analysis of data, than in the amount of data collected. 

If Artificial Intelligence can be harnessed to analyse the combined data from different sources, such as from the forestry machine sensors and aerial surveys, this could create new possibilities to recognize valuable biotopes. It can also recognise key species for biodiversity, or amounts of old, decomposing trees in certain areas. This could help the forestry machine operators in making an informed choice on where to operate.   

Artificial Intelligence has the capability to change the way we analyse and utilise data. More work is still needed to validate the results. It must be ensured that the more efficient processes provided by AI give us reliable information to support decision-making and help berry pickers find the berries.  

aUTHORS   

Annukka Pekkarinen, Communications Coordinator, Finnish Geospatial Research Institute FGI 

Harri Kaartinen, Research professor, Finnish Geospatial Research Institute FGI 

Josef Taher, Researcher, Finnish Geospatial Research Institute FGI 

Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or Directorate-General for Communications Networks, Content and Technology. Neither the European Union nor the granting authority can be held responsible for them.