One of the main objectives of the continuation phase of the FCM project is to include deep learning methodologies in the tool portfolio. Based on earlier experiences, the UNet model family approaches were selected. A vanilla UNet model as well as more advanced models with attention modules (such as earlier developed SeUNet with with a squeeze-excitation block) have been tested and further developed in the Catalonia and Norway demonstration areas. The idea is to train the convolutional models on a representative wide-area wall-to-wall reference datasets, most typically based on airborne LiDAR data. The pretrained model is then further fine-tuned at the target site using field sample plots.
The UNet mapping has some key advantages compared to the traditional machine learning approaches, such as the KNN or the Probability method. One of the key benefits is the reduced number of field plots required from the target area, when UNet-model based transfer approach is implemented. While the kNN approach typically requires hundreds of field plots from the target area, a few dozen representative plots are sufficient to fine-tune a UNet model (Ge et a. 2023). This is a great benefit for users looking to reduce their field measurement costs.
Another benefit is the contextual nature of the UNet modelling. While the kNN typically relies on spectral signatures of individual pixels, the UNet method takes into account the contextual information (i.e. the surroundings) of the pixel. This results in a more natural-looking distribution of forest variable values, with clearer distinction of adjacent forest stands. It also enables prediction of higher volume and biomass, reducing the saturation effects observed in volume and biomass predictions (Figure 1).
Figure 1. Example of Volume prediction in the Norway use case area. Colour range 0-450 m3/ha. Note the wider range of values in the UNet map on the left.
During the past months, we have developed and tested technical approaches to enable large area processing of forest variable maps with UNet models in the target areas. As sort of a proof of concept we have just finished the production of forest variable maps for the Norway use case area for the years 2017, 2019, 2021 and 2023. The area consists of 35 Sentinel-2 tiles and in the production we used Sentinel-2, Sentinel-1 and PALSAR-2 data. The newly produced maps will be used as input for the data assimilation process described in the previous blog post.
The first tests indicate that the UNet maps have significantly higher accuracy than the kNN maps. However, it is too early to make overall conclusions of the benefits and limitations of the UNet method. Full accuracy assessment and output product analysis will be conducted and reported in scientific journals in due time. The maps will also be used to study autocorrelation in the Norwegian use case area. This analysis is expected to provide valuable information on the local level accuracy of the maps (at stand or forest compartment level).
Figure 2. Example of density scatter plots in the Catalonia use case area.
Stay tuned to hear more on our UNet advancements and other findings as we learn more during the coming months.