Remote sensing-based biomass estimation of dry deciduous tropical forest using machine learning and ensemble analysis.

Autor: Singh C; Department of Environmental Science and Engineering, Indian Institute of Technology (Indian School of Mines), Dhanbad, India; Stockholm Resilience Centre, Stockholm University, Stockholm, Sweden. Electronic address: chandrakant.singh@su.se., Karan SK; Department of Environmental Science and Engineering, Indian Institute of Technology (Indian School of Mines), Dhanbad, India; Department of Energy and Technology, Swedish University of Agricultural Sciences, Uppsala, Sweden., Sardar P; Department of Environmental Science and Engineering, Indian Institute of Technology (Indian School of Mines), Dhanbad, India., Samadder SR; Department of Environmental Science and Engineering, Indian Institute of Technology (Indian School of Mines), Dhanbad, India. Electronic address: samadder@iitism.ac.in.
Jazyk: angličtina
Zdroj: Journal of environmental management [J Environ Manage] 2022 Apr 15; Vol. 308, pp. 114639. Date of Electronic Publication: 2022 Feb 09.
DOI: 10.1016/j.jenvman.2022.114639
Abstrakt: Forests play a vital role in maintaining the global carbon balance. However, globally, forest ecosystems are increasingly threatened by climate change and deforestation in recent years. Monitoring forests, specifically forest biomass is essential for tracking changes in carbon stocks and the global carbon cycle. However, developing countries lack the capacity to actively monitor forest carbon stocks, which ultimately adds uncertainties in estimating country specific contribution to the global carbon emissions. In India, authorities use field-based measurements to estimate biomass, which becomes unfeasible to implement at finer scales due to higher costs. To address this, the present study proposed a framework to monitor above-ground biomass (AGB) at finer scales using open-source satellite data. The framework integrated four machine learning (ML) techniques with field surveys and satellite data to provide continuous spatial estimates of AGB at finer resolution. The application of this framework is exemplified as a case study for a dry deciduous tropical forest in India. The results revealed that for wet season Sentinel-2 satellite data, the Random Forest (adjusted R 2  = 0.91) and Artificial Neural Network (adjusted R 2  = 0.77) ML models were better-suited for estimating AGB in the study area. For dry season satellite data, all the ML models failed to estimate AGB adequately (adjusted R 2 between -0.05 - 0.43). Ensemble analysis of ML predictions not only made the results more reliable, but also quantified spatial uncertainty in the predictions as a metric to identify its robustness.
(Copyright © 2022 Elsevier Ltd. All rights reserved.)
Databáze: MEDLINE
Externí odkaz: