Through the hard work and good preparation done over the first nine months, we have reached another exciting phase in the Forest Carbon Monitoring project. The first forest structure variable estimation test results are coming out. We have tested the estimation of variables like basal area, diameter, height and growing stock volume, as well as species identification and site type classification.
We have an ambitious testing scheme, with the aim of testing different types of algorithms (e.g. Probability, K-NN, BIOMASAR) in six testing sites around Europe and one in Peru, with different input data combinations. One key difference in the algorithms is that some of them require high numbers of field sample plots also in operational setting, while others can be run without any field reference data in an operational system. It is important to understand how much this affects the accuracy of the output products to be able to provide the user with an optimal selection of algorithms for different situations. Another interesting aspect in the testing phase is to find optimal combinations of satellite and auxiliary datasets. Optical Sentinel-2 as well as radar Sentinel-1 and PALSAR-2 data are the main datasets tested. Also ICEsat-2 space lidar and TanDEM-X radar data is tested.
The first results show that empirical methods which use field reference data in operational implementation produce more accurate estimations than purely physical models on local level. However, they can only be used in areas where representative field reference data is available, which may be a limiting factor for large area mapping (e.g. on continental level). The benefit of physical models which do not require field reference data is the ease of implementation in operational setting. So far, all the tests have been compared with pixel/plot level accuracies. Comparison of results in larger geographical areas will shed more light into the differences between the algorithms for larger area mapping.
Figure 1. Early examples of forest variable estimation with the Probability method. Top left: Sentinel-2 satellite image (RGB: SWIR 2,1; NIR; Red), Top right: Growing Stock Volume (0-350 m3/ha from white to dark green), Bottom left: Species proportion (RGB: Pine, Spruce, Broadleaf), bottom right: Scatter plot of Growing Stock Volume estimation on accuracy reference plots.
In the first tests, optical Sentinel-2 data has proven the strongest dataset individually for forest structure variable estimation, with only minor benefit from fusion with radar data. Radar data alone has resulted in somewhat lower accuracies, but with the benefit of all weather monitoring capability. The best results have been reached in our Ireland testing site, with the relative RMSE for basal area, diameter, height and growing stock volume ranging between 19% and 30% of mean, on pixel/plot level.
In other parts of Europe, the accuracies have been somewhat lower, in some cases close to 60% of mean RMSE. A major limitation, particularly for more voluminous forest, is the saturation of estimates to a certain level. From the satellite perspective, a forest with 400 m3/ha in Finland has the same spectral properties as a forest with 600 m3/ha (Figure 1). This is a well-known problem in Earth observation based forest monitoring and it has been clearly manifested also in the first tests in our project. An attempt to overcome this problem with a novel satellite derived tree height estimation approach (still a secret!) is in progress. We have high hopes that this approach would improve the accuracy further and reduce the saturation problem.
The testing will go on until June 2022, at which point we will choose the algorithms to be used in the demonstration cases. Stay tuned for more updates as we get more testing results!