Day 4

Detailed paper information

Back to list

Paper title The outcomes of the ESA MULTI ACTOR FOREST INFORMATION SERVICE (MAFIS) project about the application of L-band SAR, Sentinel-2 and LiDAR data in retrieving Above Ground Biomass by means of a Random Forest approach
  1. Matteo Picchiani GMATICS Speaker
  2. Cristina Vittucci Tor Vergata University of Rome
  3. Chiara Clementini GMATICS
  4. Marcello Maranesi GMATICS
  5. Marco Marchetti University of Molise
  6. Marco Ottaviano University of Molise
  7. Remo Bertani RDM Progetti
  8. Simone Luppi RDM Progetti
  9. Daniele Latini GEO-K srl
  10. Fabio Del Frate Tor Vergata University of Rome
Form of presentation Poster
  • A3. Biosphere
    • A3.06 Biomass monitoring
Abstract text Several studies have highlighted the saturation effects of L-Band SAR signal sensitivity at the increasing of forest density. In those cases, a direct modeling approach or an empirical regression guided by ground sampled measurements can be not effective to estimate the Above Ground Biomass (AGB) values higher than ~ 150 – 200 t/ha. Machine learning approaches are therefore proposed in recent literature to deal with these types of constrain with active (and passive) microwave monitoring of forest, by including different type of ancillary information.
In the ESA MAFIS project we have teste the feasibility of a Random Forest (RF) procedure including SAR and optical data. The strength of the RF solution consists in the possibility to include different type of Earth observation quantities in addition to the L band backscatter for characterizing the AGB. In this way the L band SAR signal is coupled with multispectral optical indexes, for limiting the saturation effects of the SAR signal, without explicitly dealing with the complex non-linearity of the combination of the input variables. Anyway this hypothesis can be effectively exploited only if a sufficient set of reference data for the AGB are available. In general, the use of in situ measurements sampled on tens to hundreds grounds plots cannot be sufficient to proper finalize the training phase of a data drive algorithm. In the MAFIS project we tried to overcame such limitation by exploiting recent aerial LiDAR data made available for the Veneto Region over the alpine areas of Lorenzago di Cadore and Bosco del Cansiglio (Noth-East of Italy). Those areas were affected by the Vaia storm, occurred from the 26th to the 30th October 2018. This event caused a dramatic loss of forest area in different Italian regions due to strong winds that pulled down a massive quantitative of trees. Regione Veneto acquired a large set of aerial LiDAR data after the Vaia storm for mapping the extension of the affected areas. This quite unique dataset represent a good opportunity to evaluate the effectiveness of a Random Forest approach for the AGB retrieval by means of the fusion of L-band SAR data and multispectral data. In fact the forest areas acquired during the flights spans for several tens of hectares and provides thousands of training examples of intact areas over the forest AGB can be derived from the LiDAR measurements of tree’s height. In particular, the LiDAR data have been processed to derive the Digital Terrain model (DTM) and the Digital Surface model (DSM). These latter have been used to derive the tree height layer over the forest considered areas. Finally, a corrected version (fitted to several local data acquired during the MAFIS project in situ survey) of dendrometric tables of the second Italian National Forest Inventory (INFC), which define volume estimation equations adapted to the different forest species, have been applied to the most common tree species of the considered Alp regions, i.e. Fagus, Abies Alba and Larix Decidua, for computing the LiDAR based AGB layer, which ranges between ~200 and ~1000 m3/ha over the analysed regions. The latter has been the divided in the training and test set, used respectively to train the RF model and to test its performances.
The input data to the RF model are the HH and HV backscattering coefficients, extracted from ascending and descending SAR ALOS-2 PALSAR-2 L1.1 products, and multispectral reflectance (in VIS, NIR an SWIR), extracted from Sentinel-2 L2A products. Both the ALOS-2 and the Sentinel-2 data have been collected on specific dates comparable with the time range of the aerial LiDAR acquisitions.
The results of the trained RF model evaluated over the independent test set shown very encouraging results, with a correlation coefficient higher than 70% and reporting very coherent behavior of the spatial patterns of AGB within the mountain landscape. Finally, the insurgence of the saturation effects is registered for a threshold of about 900 m3/ha.