Improving LULUCF carbon emissions/removals estimates for Flanders, Belgium, through high-resolution prediction of land use based on a machine learning approach

The quality of national and regional estimates of carbon emissions/removals under the LULUCF sector directly depends on the quality of the used input data for land use and land use changes, and their associated carbon stocks and emission/removal factors. The increasingly strict European regulations for greenhouse gas (GHG) emission reporting and the growing importance of the LULUCF sector in climate action plans and policies provide clear incentives to strive for continuous improvement of LULUCF datasets, including their spatial and temporal resolution.

The region of Flanders, Belgium, is characterized by relatively heterogeneous land use, which is monitored at a high spatial resolution resulting in the triennial production of a detailed land use map (18 categories, raster map with cell size of 10 by 10 meters). However, the estimation of LULUCF carbon emissions/removals currently relies on a different, less detailed land use dataset (5 categories, regular grid of 6799 points each representing an area of approximately 2000 by 1000 meters). The use of the latter dataset is motivated by the adoption of a similar approach over the three regions of Belgium, to guarantee the consistent integration of regional carbon emissions/removals estimates into the national GHG inventory. Apart from the general limitations of a sample-based dataset, this LULUCF land use dataset provides insufficient detail to grasp the effect of LULUCF-related policies and measures, undermining the use of derived carbon emissions/removals estimates for policy evaluation and development at the level of Flanders.

To overcome the issues related with spatial (and temporal) resolution of the current LULUCF land use dataset, we tested a machine learning approach to integrate four more detailed data sources available for Flanders in order to predict the corresponding LULUCF land use at a resolution of 10 meters. More specifically, we used the land use file, the land use map, the land cover map, and the dataset of registered agricultural parcels as input data in a multinomial logistic regression, considering a search neighbourhood with a 10-m radius around an original LULUCF land use data point. The model was created based on input data for two years, namely 2012 and 2015. Of the original LULUCF dataset, 70% of the data points was used for training of the model, leaving 30% for validation. In a first test, a prediction accuracy of approximately 90% was achieved. After manual correction of the original LULUCF dataset, the accuracy improved to 94%.

Although the approach proved successful for the prediction of the LULUCF land use for the individual years considered, less satisfying results were found when using the predictions to derive the land use change between these years: a land use change was estimated to occur at 8% of the total area of Flanders, while this was only 2% and 4% based on the original dataset before and after manual correction. Considering the major significance of the land use change area in the estimation of carbon emissions/removals, further research is required to adjust the methodology in order to guarantee the prediction of a consistent land use time series.