An indirect approach to predict deadwood biomass in forests of Ukrainian Polissya using Landsat images and terrestrial data

Maksym Matsala,Andrii Bilous,Viktor Myroniuk,Roman Zadorozhniuk,Petro Diachuk,Andrii Terentiev

doi:10.2478/fsmu-2020-0018

Abstract

Abstract Spatially explicit and consistent mapping of forest biomass is one of the key tasks towards full and appropriate accounting of carbon budgets and productivity potentials at different scales. Landsat imagery coupled with terrestrial-based data and processed using modern machine learning techniques is a suitable data source for mapping of forest components such as deadwood. Using relationships between deadwood biomass and growing stock volume, here we indirectly map this ecosystem compartment within the study area in northern Ukraine. Several machine learning techniques were applied: Random Forest (RF) for the land cover and tree species classification task, k-Nearest Neighbours (k-NN) and Gradient Boosting Machines (GBM) for the deadwood imputation purpose. Land cover (81.9%) and tree species classification (78.9%) were performed with a relatively high level of overall accuracy. Outputs of deadwood biomass mapping using k-NN and GBM matched quite well (8.4 ± 2.3 t·ha−1 (17% of the mean) vs. 8.1 ± 1.7 t·ha−1 (16% of the mean), respectively mean ± SD deadwood biomass stock), indicating a strong potential of ensemble boosters to predict forest biomass in a spatially explicit manner. The main challenges met in the study were related to the limitations of available ground-based data, thus showing the need for national statistical inventory implications in Ukraine.

Highlights

Explicit estimation of forest structural compartments has become a main target of studies and forest monitoring during the last decades (Lim et al, 2003; Zimble et al, 2003; Zald et al, 2016)
Since the allometric models of deadwood biomass are consistent with the parameters of a growing forest stand, we directly model this compartment instead of linking to growing stock volume (GSV) values (Pflugmacher et al, 2012)
Land cover types “water”, “forest” and “croplands” on the validation dataset were interpreted with a high level of user accuracy (93.1%, 92.2%, and 90.0%, respectively), while the largest misclassifications were met for the types “bog”, “other” and “shrublands” (Table 2)

Summary

Introduction

Explicit estimation of forest structural compartments has become a main target of studies and forest monitoring during the last decades (Lim et al, 2003; Zimble et al, 2003; Zald et al, 2016). Deadwood as the key legacy component of forest ecosystems reflecting natural dynamics processes in the stand structure, is increasingly considered in conservation and management policies throughout the world (Franklin et al, 2002; Seidl et al, 2014). New temporally continuous and spatially explicit approaches have significantly empowered a robust estimation of deadwood (Gonzalez et al, 2018). Temporal segmentation of Landsat seasonal mosaics allows an identification of natural disturbances which are strictly linked to deadwood accumulation (Cohen et al, 2010; Kennedy et al, 2010). Significant improvement of deadwood estimation has been achieved with rising application of terrestrial-based and airborne LiDAR techniques coupled with empirical field-based and satellite-based data (Gonzalez et al, 2018)

Objectives

Methods

Results

Discussion

Conclusion