Publications

Fararoda, R; Reddy, RS; Rajashekar, G; Chand, TRK; Jha, CS; Dadhwal, VK (2021). Improving forest above ground biomass estimates over Indian forests using multi source data sets with machine learning algorithm. ECOLOGICAL INFORMATICS, 65, 101392.

Abstract
Accurate estimates of spatial above ground biomass (AGB) in tropical forests are important in understanding the global carbon cycle. Microwave and optical remote sensing datasets have been used extensively for AGB estimation, but their uses are restricted due to saturation in high biomass region. To overcome saturation issues of single sensor based models, the current paper uses a non-parametric Random Forest based approach to spatially estimate biomass over Indian forests using field inventory data in combination with optical (MODIS) and Microwave (L-band ALOS-PALSAR) images along with other bio-climatic parameters (e.g., rainfall, temperature) which significantly influence the biomass accumulation in an ecosystem. Plot level biomass estimates for 6678 sample plots (0.1 ha size), inventoried as part of robustly designed National Forest Inventory (NFI), are computed using volumetric equations, wood density and biomass expansion factors. Spatial above ground biomass estimates were generated using random forest model over two scenarios. Firstly, a single nation-wide model using all the available plot data and secondly, physiographic zone (14 zones over India) wise models with plot data over respective zones. AGB stored in Indian forest is estimated as 7952.3 million tonnes (Carbon equivalent: 3737.58 TgC) with a root mean square error (RMSE) of 31.2% using national level model. Physiographic zone level models estimated the country's biomass as 7597.45 million tonnes (Carbon equivalent: 3570.8 TgC). The above ground biomass estimates from our study indicates that the estimation error in physiographic zone model varies from 25.24% to 54.15% depending upon the sample size and biomass range. We observed that L-band microwave saturated in 140-160 Mg ha-1 range and currently available microwave wavelengths alone is not sufficient to predict entire range of biomass over Indian forests. Inclusion of multisource data using random forest regression model increase the saturation range to 350 Mg ha-1 which is a significant improvement as 94.7% of Indian forests are covered in this range. Model estimation error reduces to 25.6% in AGB range up to 350 Mg ha-1.

DOI:
10.1016/j.ecoinf.2021.101392

ISSN:
1574-9541