Estimating Aboveground Biomass of Two Different Forest Types in Myanmar from Sentinel-2 Data with Machine Learning and Geostatistical Algorithms

Abstract

The accurate estimation of spatially explicit forest aboveground biomass (AGB) provides an essential basis for sustainable forest management and carbon sequestration accounting, especially in Myanmar, where there is a lack of data for forest conservation due to operational limitations. This study mapped the forest AGB using Sentinel-2 (S-2) images and Shuttle Radar Topographic Mission (SRTM) based on random forest (RF), stochastic gradient boosting (SGB) and Kriging algorithms in two forest reserves (Namhton and Yinmar) in Myanmar, and compared their performance against AGB measured by the traditional methods. Specifically, a suite of forest sample plots were deployed in the two forest reserves, and forest attributes were measured to calculate the plot-level AGB based on allometric equations. The spectral bands, vegetation indices (VIs) and textures derived from processed S-2 data and topographic parameters from SRTM were utilized to statistically link with field-based AGB by implementing random forest (RF) and stochastic gradient boosting (SGB) algorithms. Followed by an evaluation of the algorithmic performances, RF-based Kriging (RFK) models were employed to determine the spatial distribution of AGB as an improvement of accuracy against RF models. The study’s results showed that textural measures produced from wavelet analysis (WA) and vegetation indices (VIs) from Sentinel-2 were the strongest predictors for evergreen forest reserve (Namhton) AGB prediction and spectral bands and vegetation indices (VIs) showed the highest sensitivity to the deciduous forest reserve (Yinmar) AGB prediction. The fitted models were RF-based ordinary Kriging (RFOK) for Namhton forest reserve and RF-based co-Kriging (RFCK) for Yinmar forest reserve because their respective R2, whilst the RMSE values were validated as 0.47 and 24.91 AGB t/ha and 0.52 and 34.72 AGB t/ha, respectively. The proposed random forest Kriging framework provides robust AGB maps, which are essential to estimate the carbon sequestration potential in the context of REDD+. From this particular study, we suggest that the protection/disturbance status of forests affects AGB values directly in the study area; thus, community-participated or engaged forest utilization and conservation initiatives are recommended to promote sustainable forest management.

ReUse: REgressive Unet for Carbon Storage and above Ground Biomass Estimation

“…In this paper, we introduce a Regressive UNet trained on the public above-ground biomass (AGB) data from the European Space Agency's Climate Change Initiative Biomass project as ground truth to estimate the carbon sequestration capacity of any portion of land on Earth using Sentinel-2 images, comparing its performance against two literature proposals [10,14] on their respective study areas. Section 3.1 introduces the proposed approach, describing ideas and motivations.…”

“…In [12], Sentinel-2 performance was evaluated for a buffer zone community forest in Parsa National Park, Nepal, using field-based AGB as a dependent variable, as well as spectral band values and spectral-derived vegetation indices as independent variables in the Random Forest (RF) algorithm; in this study, no features were extracted from the spatial dimensions, but indicators were only extracted from the spectral dimension of the input tensor. In [14] spectral bands, vegetation indices (VIs) and texture variables derived from processed S-2 data and topographic parameters were utilized to statistically link with field-based AGB by implementing random forest (RF) and stochastic gradient boosting (SGB) algorithms. The grey level co-occurrence matrix (GLCM) method [16] and wavelet decomposition were applied using the first principal component of the Sentinel-2 multispectral tensor.…”

“…To the best of our knowledge, this is the first time such an approach has been used to estimate AGB. The advantage of this approach over classical approaches, such as [10] and [14], is that the extraction of both spatial and spectral features is incorporated into the same network in an end-to-end paradigm. A patch-wise approach was followed as in [18] so that each Sentinel-2 input and AGB-raster output were divided into many non-overlapping patches of 16*16.…”

“…ReUse was compared with Random Forest [19] by taking the feature extracted as in [10] and in [14] as a reference, in which indicators are extracted not only considering the spectral component but also taking spatial information into account. Note that [14] also uses topographical parameters such as altitude as a variable; however, in this work, only Sentinel-2 data were used as input for the model comparison.…”

“…Typically in the literature, to estimate above-ground biomass (AGB) and the related carbon stock, field surveys are carried out, and a classical machine learning model is trained on sentinel-2, and its extracted features as input and AGB's collected field measurements as output [10] [11] [12] [13] [14]. Instead, continuing without field measurements might be an optimal solution to monitor AGB and carbon storage globally and on time, even downstream of natural disasters such as large fires, which cause deforestation.…”