Remote sensing based estimation of forest biophysical variables using machine learning algorithm

Leaf Area Index (LAI), Fraction of Intercepted Photosynthetically Active Radiation (fIPAR) and forest Aboveground Biomass (AGB) are important regulatory parameters for several functions of the forest canopy. An accurate information about the spatial variability of these biophysical variables is vital to capture the variability in estimates of gross primary productivity, carbon exchange and microclimate in terrestrial ecosystems. The present study aims at developing predictive models for generating spatial distribution of LAI, fIPAR and AGB by integrating remote sensing imagery and field data using random forest (RF) regression algorithm. The study was carried out in a tropical moist deciduous forest of Uttarakhand, India. Various spectral and texture variables were derived using Sentinel-2 data of 10 April 2017. In-situ measurements of LAI, incident Photosynthetically Active Radiation (PAR) above canopy (I_o), below canopy (I), and diameter at breast height (dbh) were taken. fIPAR and AGB were calculated. RF regression algorithm was used to optimize the variables to select the best predictor variables. Three models, using only spectral variables, only texture variables and both spectral and texture variables were tested. For all three biophysical variables, the models using both spectral and texture variables gave better results. The best predictor variables were used to map the spatial distribution of LAI, fIPAR and AGB. On validation, the models were able to predict LAI with R^2=0.83, %RMSE = 13.25%, fIPAR with R^2=0.87, %RMSE = 13.24%, and AGB with R^2=0.85, %RMSE = 12.17%. The estimated biophysical parameters showed high interdependence (LAI-fIPAR R2= 0.71, LAI-AGB R^2=0.75 and fIPAR-AGB R^2= 0.74). The results showed that RF can be effectively applied to predict the spatial distribution of forest biophysical variables like LAI, fIPAR and AGB with adequate accuracy.