Even a machine can train to be a forester
The occurrence of the spruce bark beetle (Ips typographus) in Central Europe has been increasing over the past three decades. The most recent bark beetle outbreaks in the Czechia (since 2015) even exceed the frequency and severity of outbreaks observed over the previous two decades. It is therefore time to act, as the possibility remains that we can stop, or at least slow, the death of our forests through the early detection of bark beetle activity.
One way to detect tree infestations at an early stage is through field surveys. However, this is an extremely time-consuming activity and it is not capable of covering a large area. The use of remote sensing offers a considerably more comprehensive solution. Here, however, we often encounter inadequate image resolution and the impossibility of detecting the specific crowns of infested trees. Moreover, knowledge of the interactions between the biochemical and structural properties of the tree and the electromagnetic signals detected by the sensor is needed.
A promising tool for identifying infected trees is image classification using machine learning algorithms (a subfield of artificial intelligence where a computer system ‘learns’ from the presented patterns and their key features). One way to process multispectral images with very high resolution is to use the convolutional neural network (CNN) method. The hyDRONE team, composed of physical geography experts and drone enthusiasts, decided to use this modern approach to classify individual tree crowns. They situated their research in the Klánovice Forest in Prague which has been suffering from drought and associated massive outbreaks of bark beetles since 2015. This makes it an ideal pilot site. Firstly, they divided the crowns of the individual trees into four main categories: pines, longer infested trees whose needles turn yellow, trees under green attack, and non-infested trees. In conjunction with drone imaging, a field survey was conducted.
The next step was to perform an image analysis using convolutional neural networks (CNN). A CNN consists of a chain of convolutional and pooling layers terminated by a classification layer of neurons. Individual mini images of the crowns enter the network. In the convolutional layer, typical features of the crowns for a given category are highlighted by multiplying the pixels with a moving filter called a kernel. The pooling layer reduces the image size and condenses the information by selecting the maximum or calculating the average from a defined array of, for example, 3 × 3 pixels. This reveals the dominant features of the input data hidden in the matrices that enter the classifier. Similar to biological processes, CNNs use imaginary neurons to convey information and make subsequent responses to a given stimulus. The advantage of a CNN (unlike other machine learning methods) is that it ‘searches’ the input images for key features (patterns) for category recognition on its own. There is no need to supply these patterns to the network through image transformations.
The authors analysed images of variously damaged forests using three different convolutional neural network models and the widely used ‘random forest’ model. CNN models using only RGB bands were found to be the most suitable because the optimal spectral separability of the classes was in the blue and red bands. There were only a few misclassifications between the two categories of spruce trees at different stages of infestation.
Early detection of bark beetle infested trees is essential to stop the spread of bark beetle and prevent further damage to our forests. In a recent study, a team of experts has revealed new ways to use convolutional neural networks to detect these beetles. The reported models can be used not only for basic identification of infested trees, but also for monitoring the evolution of bark beetle disturbance. The authors will further refine the models using more training samples with repetition under different seasonal conditions.
Kateřina Fraindová
Minařík, R., Langhammer, J., Lendzioch, T. 2021. Detection of Bark Beetle Disturbance at Tree Level Using UAS Multispectral Imagery and Deep Learning. Remote sensing 13, 4768. https://doi.org/10.3390/rs13234768
Document Actions