SENTREE - Norwegian tree species classification
Project type: Applied Research
Region: Mid-North of Norway
Duration: 1 year (October 2021 - October 2022)
For forest management the availability of complete and up-to-date forest inventories is essential, with one of the most important parameters being the volumetric tree species distribution. Unfortunately, tree species mapping in Norwegian production forests is a time-consuming and largely manual process, leading to forest inventories that are often incomplete and/or outdated. Indeed, the determination of the tree species distribution is currently performed by a forestry expert, mainly by visual interpretation of aerial imagery and in some cases lidar data. High resolution aerial imagery is available, however campaigns are expensive and therefore infrequent. Satellite imagery, on the other hand, provides dense time series, but at a much lower resolution. The primary goal of the SENTREE project is to automate the classification of Norway spruce, Scots pine and Birch using semantic segmentation of aerial and Sentinel-2 imagery.
There are three main innovative parts in this project:
1) The combination of aerial imagery and sentinel-2 data, aiming to exploit the high spatial resolution of the aerial imagery and high spectral and temporal resolution of the Sentinel-2 imagery in a single network.
2) The application of noise-robust training and noise detection using custom loss functions and AUM respectively. The tree species training labels are known to contain some error, handling this noise is a major challenge.
3) The application of semi-supervised training schemes by using pseudo labelling and consistency loss. The availability of labelled tree species data is limited, semi-supervised learning is exploited to train and improve performance over regions where labelled data is not available.
Our initial models show promising results with overall good correspondence between the predicted tree species by the model and reference data. The models are able to exploit textural information in the aerial imagery and spectral/temporal information from the Sentinel-2 data to discriminate between the different species. The example below shows a close-up of the tree species classification.
Figure 1: Norway spruce (dark-green), Scots Pine (brown) and Deciduous (light-green).
In addition to classifying tree species for new regions and filling empty stand data with a tree species distribution, our models could also be exploited to identify mislabelled stands in areas where a forestry inventory is already present (see red stands below).
Figure 2: Mislabelled stands (red).
Interested? We could make something cool for you too, just contact us!