Day 4

Detailed paper information

Back to list

Paper title Modelling of forest aboveground biomass from satellite optical and radar observations
  1. Barbora Navratilova Global Change Research Institute CAS Speaker
  2. Olga Brovkina Global Change Research Institute of the Czech Academy of Sciences Speaker
  3. Jan Novotny CzechGlobe
Form of presentation Poster
  • A3. Biosphere
    • A3.06 Biomass monitoring
Abstract text Aboveground forest biomass (AGB) accounts for between 70% to 90% of total forest biomass estimates, which are the central basis for carbon inventories. Estimation of forest aboveground biomass (AGB) is critical for regional forestry and sustainable forest management. Remote sensing (RS) data and methods offer opportunities of AGB broad-scale assessments providing data over large areas at a fraction of the cost with access to inaccessible places. Optical RS provides good alternative to biomass estimation through field sampling due to its global coverage, repetitiveness and cost-effectiveness. Radar RS has gained prominence for AGB estimation in recent years due to its cloud penetration ability as well as detailed vegetation structural information.
In this study, the potential of C-band SAR data from Sentinel-1, L-band SAR data from ALOS PALSAR, multispectral data from Sentinel-2 instruments and machine learning algorithms were evaluated for the estimation of AGB in a mountainous mixed forest in the Eastern part of the Czech Republic. The response variable was AGB (Mg/ha) estimated from normalized digital surface model nDSM (Forest Management Institute, and field measurements (R2=0.84, nRMSE = 10%). The following cases of predictors were considered for AGB modelling: (1) Sentinel-1, Sentinel-2 and ALOS PALSAR, (2) Sentinel-1 and Sentinel-2, (3) ALOS PALSAR and Sentinel-2. SAR data were used with VV and VH polarizations. Normalized difference vegetation index NDVI, tasselled cap transformation TC (greenness, brightness and wetness) and disturbance index DI were calculated from multispectral Sentinel-2 data and together with single spectral bands were used as predictors. The modeling was performed with several machine-learning algorithms including, neural network, adaptive boosting and random decision forest. The AGB models were developed for coniferous, deciduous and mixed types of forest. AGB estimates for deciduous forest stands generally showed a weaker predictive capacity for all models, than AGB estimates for coniferous. The models with Sentinel-1 and Sentinel-2 predictors (case 2) had the weaker estimates comparing with models using ALOS PALSAR predictors (cases 1 and 3). The best model performance was achieved with the random decision forest algorithm and predictors derived from three sources of satellite data, Sentinel-1, Sentinel-2 and ALOS PALSAR. The proposed methodology can be applicable for Central European forest AGB mapping in large areas using the satellite optical and radar data.

Keywords: machine learning, ALOS PALSAR, Sentinel, forest productivity.

Acknowledgment: The study was supported by the Ministry of Agriculture of the Czech Republic, grant number QK1910150.

Cairns et al. 2018. Root biomass allocation in the world’s upland forests. Oecologia 1997, 111, 1–11.
Chen et al. 2018. Estimation of forest above-ground biomass by geographically weighted regression and machine learning with Sentinel imagery. Forests, 9.
Pedregosa et al. 2011. Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research, 12.
Perko et al. 2011. Forest assessment using high resolution SAR data in X-band. Remote Sens., 3, 792-815.