Multi-sensor remote sensing and machine learning for aboveground biomass mapping in Vietnam’s Melaleuca wetlands: A review

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025

www.rsisinternational.org Page 473

Multi-sensor remote sensing and machine learning for aboveground
biomass mapping in Vietnam’s Melaleuca wetlands: A review

Nguyen Thi Ngoc Tam1, Nguyen Vo Chau Ngan2*

1Master’s Student, Climate Change and Delta Management Program (K30 Cohort), Can Tho University,
Vietnam

2College of Environment and Natural Resources, Can Tho University, Vietnam

*Corresponding Author

DOI: https://doi.org/10.51584/IJRIAS.2025.1010000035

Received: 30 Sep 2025; Accepted: 07 Oct 2025; Published: 01 November 2025

ABSTRACT

Accurate mapping of aboveground biomass in tropical peatland forests remains challenging due to the
complexity of vegetation structure, hydrological regimes, and data heterogeneity across sensors. This review
synthesizes multi‑sensor remote sensing and machine‑learning approaches for aboveground biomass estimation
in Vietnam’s Melaleuca wetlands, aiming to establish a standardized framework of terminology, metrics, and
environmental covariates for future research and applications. By harmonizing key indicators such as canopy
height, texture, soil-hydro-geomorphological variables, and validation metrics (R², RMSE), the framework
enhances reproducibility, comparability, and data integration across scales. The study further consolidates a
practical roadmap encompassing data acquisition, feature engineering, modeling, and validation stages -
culminating in uncertainty‑aware biomass mapping that bridges research and operational implementation.
Beyond synthesizing existing studies, this work provides actionable guidance for open‑access workflows and
policy‑oriented applications in carbon accounting and wetland restoration. The proposed standardized approach
thus supports both scientific and managerial communities in advancing sustainable management of Vietnam’s
Melaleuca peat ecosystems and will help standardize future aboveground biomass mapping across Southeast
Asian wetlands.

Keywords: L-band SAR, LiDAR and GEDI, Melaleuca aboveground biomass, Multi-sensor data fusion, Spatial
and spatio-temporal cross-validation

INTRODUCTION

Melaleuca-dominated wetlands are globally important carbon reservoirs and biodiversity refuges, yet their
aboveground biomass (AGB) remains challenging to map reliably at scale. Peat accumulation, acid-sulfate soils,
strongly coupled hydrology and microtopography, and frequent radar/optical saturation in dense stands
complicate both field estimation and remote-sensing (RS) inference, producing spatially heterogeneous
allometry and sensor responses that impede model transferability. Field and remote studies have documented
substantial variation in AGB drivers and signal behavior across these substrates, underscoring the need for
tailored approaches (Huy et al., 2016; Kappas, 2020; Nam et al., 2016; Tran et al., 2015; Zadbagher et al., 2024).

Recent advances in multi-sensor fusion (light detection and ranging [LiDAR]/ global ecosystem dynamics
investigation [GEDI], L-band synthetic aperture radar (SAR), Sentinel-1/2), machine learning, and
multi-temporal analysis show promise for improving accuracy, but reported performance varies widely and is
sensitive to validation strategy, sensor choice, and environmental covariates (Balestra et al., 2024; Musthafa &
Singh, 2022; Nguyen et al., 2024; Zhang et al., 2019, 2020). Moreover, many studies report optimistic accuracies
when spatial autocorrelation is not properly accounted for; best practices now emphasize spatial or
spatio-temporal blocking for robust out-of-sample assessment (Roberts et al., 2017; Valavi et al., 2019). At the
same time, localized allometry (diameter at breast height [DBH] - height - wood density) and site-specific

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025

www.rsisinternational.org Page 474

predictors (hydroperiod, peat depth, salinity, microtopography) critically influence both AGB and sensor signals,
and must be integrated into modelling workflows to achieve transferable maps (Ngo et al., 2023; Tran et al.,
2015).

This review synthesizes these developments with the explicit aim of informing robust AGB mapping in
Melaleuca wetlands. We (i) summarize methodological accuracy and observed performance ranges and their
drivers, (ii) evaluate multi-sensor and temporal data integration strategies, (iii) discuss computational and
practical modelling considerations, and (iv) identify the hydrology - geomorphology - soil covariates most
important for transferability. Throughout, we adopt consistent terminology (e.g., L-band SAR, LiDAR, GEDI,
Sentinel-1/2) and report performance metrics using the coefficient of determination (R2) and root mean square
error (RMSE) (Mg·ha-1) for comparability. The review concludes with a concise practical roadmap for
scaffolded, uncertainty-aware mapping suited to Melaleuca landscapes.

Remote Sensing-Based Aboveground Biomass Mapping: Methods, Data Integration, Validation, And
Scaling

Estimating AGB from RS underpins carbon accounting, REDD+, and long-term forest monitoring. In Vietnam
where Melaleuca cajuputi (cajeput) forests are widespread across the Mekong Delta, models must be both
accurate and scalable, yet sensitive to site-specific conditions (peat soils, acid sulfate soils, and fluctuating
hydrology). Recent literature shows a shift from traditional regressions toward ML and multi-sensor data fusion
(optical - SAR - LiDAR/UAV), coupled with stricter spatial-temporal validation schemes to avoid optimistic
accuracy assessments. This section provides a review of methodological accuracy, data integration,
computational complexity, and spatio-temporal resolution, and discusses their implications for Melaleuca forests
in Vietnam.

METHODOLOGICAL ACCURACY

Reported model performance for RS-based AGB is commonly summarized with R2 and RMSE. Across studies,
observed R2 values vary widely (approx. 0.59 - 0.95), with RMSE dependent on forest type, sensor combination,
and modelling strategy; for comparability this review reports RMSE in Mg·ha-1 where possible (Nguyen et al.,
2024; Zhang et al., 2019, 2020). In general, models that integrate structural information (LiDAR/GEDI) with
spectral and radar predictors and that use advanced learning architectures report the highest fits (deep-learning
examples reaching R2 ≈ 0.93 - 0.95 in some settings), while single-source optical models more frequently
occupy the lower end of the observed range. Structurally complex or radar/optically saturated systems (e.g.,
tropical peat swamp forests) often yield substantially lower test R2 (≈ 0.21 - 0.70), reflecting signal saturation
and heterogeneous allometry (Zadbagher et al., 2024).

Machine learning approaches (random forest [RF], support vector regression, gradient-boosted trees, and neural
networks) typically reduce prediction error relative to simple linear or multiple regression baselines, particularly
when spectral and structural predictors are combined; ensemble or stacking strategies further improve robustness
in many comparative studies (Chen et al., 2023; Khan et al., 2024; Nguyen et al., 2024; Zhang et al., 2020).

A critical caveat is validation strategy. Random k-fold cross-validation (CV) that ignores spatial (and temporal)
autocorrelation routinely produces optimistic accuracy estimates. To obtain realistic out-of-sample performance
and to assess transferability across hydrological zones or management units, studies should use spatial or
spatio-temporal blocking (with block sizes informed by empirical autocorrelation) or leave-one-region-out tests
(Roberts et al., 2017; Valavi et al., 2019). Authors are also encouraged to report multiple complementary metrics
(e.g., R2, RMSE in Mg·ha-1, MAE, bias) and to quantify predictive uncertainty (e.g., bootstrap, quantile
estimates, or prediction intervals) so that users can judge both central tendency and spread.

We recommend reporting both central tendency and uncertainty - providing R2 and RMSE (Mg·ha-1) together
with prediction intervals or quantiles - prioritizing multi-sensor, structurally informed predictor sets
(LiDAR/GEDI with SAR and optical) in saturation-prone or structurally complex stands; assessing
transferability using spatial or spatio-temporal blocking (or leave-one-region-out) to avoid optimistic bias; and
adopting tree-based ensembles (RF, extreme gradient boosting [XGBoost]/light gradient boosting machine

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025

www.rsisinternational.org Page 475

[LightGBM]) as robust baselines, reserving complex deep architectures for cases with substantial structural
supervision and large training samples.

Table 1 provides a concise summary of the objectives, data sources, model families, validation strategies, and
metrics referenced throughout Section 2.

Data integration level

A growing body of evidence shows that integrating multiple sensors and field data materially improves AGB
estimation by mitigating sensor-specific limitations (e.g., optical/radar saturation) and by recovering canopy
structure. In practice, structural sources (LiDAR or GEDI) combined with L-band SAR and high-resolution
optical imagery (e.g., Sentinel-2) generally outperform single-source configurations, particularly in
high-biomass or heterogeneous wetlands (Balestra et al., 2024; Musthafa & Singh, 2022; Vafaei et al., 2018;
Wang et al., 2023).

Field plots and ecologically relevant covariates (topography, peat depth, soils, hydrological metrics) remain
essential for calibration and for improving model generalization across substrate and management gradients.
Local allometry (DBH - height - wood density) provides mechanistic anchors for translating structural
predictions into AGB and reduces bias when substrate properties vary (Tran et al., 2015; Zadbagher et al., 2024).

Temporal depth further strengthens inference: multi-epoch LiDAR or GEDI combined with continuous
optical/SAR time series captures biomass trajectories (decline, recovery) and improves long-term monitoring
and change detection (Loh et al., 2022; Musthafa & Singh, 2022; Naik et al., 2021).

Table 1. Summary comparison of workflow stages for AGB mapping

Stage Key
recommendation

Typical
sensors/data

Recommended
models

Validation Metric
s

Data
acquisitio

Use structural
scaffolds +
wall-to-wall

sensors; include
field plots and env

covariates

LiDAR
(UAV/airborne) or

GEDI;
Sentinel-1/2;

L-band SAR; field
plots;

hydrological/topo/
soil layers

N/A (data stage) N/A Data
vol high
(depend

s on
LiDAR

)

Feature
engineeri
ng /fusion

Late feature fusion
of structure +

spectral + SAR;
include peat/soil/

hydrological
covariates

Canopy height
model metrics,

height percentiles,
spectral indices,
SAR backscatter
and polarimetry,

peat depth,
hydroperiod

RF/XGBoost/LightG
BM → ensemble
stacking → deep
learning (DL) (if
abundant labels)

N/A (see
modelling)

mediu
m

Modellin
g

Start with
tree-based

baselines; progress
to DL only with
dense structural

supervision

As above RF/XGBoost/LightG
BM (baseline);

Stacking ensembles;
DL (CNNs) when

LiDAR/ GEDI
supervision present

Use
spatial/spatio-temp
oral blocking for
hyperparameter

tuning

Baselin
e: low -
mediu
m; DL:

high

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025

www.rsisinternational.org Page 476

Validatio
n and

transfer
-ability

Use
spatial/spatio-temp
oral block CV or

leave-one-region-o
ut; report stratified

errors

Validation samples
from LiDAR

strips/ independent
plots

N/A Spatial block CV
(block size

informed by
autocorrelation);

report out-of-region
tests

low -
mediu

Metrics
and

uncertaint
y

Report central
tendency + spread;

stratify by
substrate/hydro

class

R2, RMSE
(Mg·ha-1), MAE,

bias; show
prediction

intervals/quantiles

N/A Report metrics for
each validation fold

and by strata

low

Operation
al notes

Document
preprocessing,
seeds, versions;
present stratified

error maps

Radiometric/terrai
n correction,

co-registration,
speckle filtering

N/A

Computational complexity

Computational cost and model complexity are central considerations when selecting methods for AGB mapping
because gains in predictive accuracy frequently entail higher requirements for compute, storage, and labeled
structural supervision. Deep architectures (e.g., convolutional neural networks, autoencoders) and ensemble
stacking/boosting typically deliver superior accuracy and scalability for large, multi-sensor datasets, but they
demand substantial graphics processing unit (GPU)/central processing unit (CPU) resources, careful
hyperparameter tuning, and rigorous data preprocessing, especially for regional-scale, multi-source inputs
(Khan et al., 2024; Zhang et al., 2019).

By contrast, traditional regression and simple allometric models remain computationally lightweight and
interpretable but often generalize poorly in settings affected by spectral/ radar saturation or high structural
heterogeneity. Tree-based methods (RF, XGBoost/LightGBM) offer a pragmatic middle ground: they are
computationally efficient, robust to heterogeneous predictors, and serve as strong baselines when training data or
computational budgets are limited (Chen et al., 2023; Zhang et al., 2020).

From a practical data-science perspective, we recommend the following conventions: (i) use spatial or
spatio-temporal block CV (with block sizes guided by empirical spatial autocorrelation) for hyperparameter
selection to avoid leakage and inflated performance estimates; (ii) standardize multi-source preprocessing
(radiometric and terrain correction, co-registration, speckle filtering) and document these steps; and (iii) adopt
late feature fusion with tree-based learners as robust baselines, moving to compact DL solutions only when
abundant structural supervision (e.g., LiDAR/GEDI scaffolds) and large, diverse training samples are available
(Nguyen et al., 2024; Roberts et al., 2017; Valavi et al., 2019).

When sample sizes or computational resources are constrained, RF, XGBoost, or LightGBM should serve as the
default baselines. Deep learning should be reserved for settings with dense labels and structural scaffolds (e.g.,
LiDAR/GEDI), with training and inference compute quantified and reported. Hyperparameters should be
selected using spatial or spatio-temporal blocking to obtain realistic estimates of transferability. Finally,
preprocessing pipelines should be fully documented, with reproducible settings (random seeds, software
versions) explicitly stated.

Temporal and spatial resolution

Spatial and temporal resolution jointly determine the suitability of sensors and modelling strategies for AGB
estimation: fine spatial detail captures structure at plot and stand scales, whereas temporal depth enables
monitoring of dynamics and disturbance-driven trajectories.

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025

www.rsisinternational.org Page 477

At the plot and stand scale, UAV or airborne LiDAR delivers high-fidelity three-dimensional structure (e.g.,
CHM metrics and height percentiles) that correlate strongly with AGB and are especially valuable when fused
with high-resolution optical imagery (Wang et al., 2023; Yan et al., 2024). These structural data are ideal for
deriving local allometry and for calibrating models that require detailed canopy scaffolding.

For regional, wall-to-wall mapping the recommended approach is to use LiDAR/GEDI strips or footprints as
calibration scaffolds and extrapolate with satellite imagery and SAR (Sentinel-1/2 and L-band where available).
This scaffold-and-extrapolate strategy typically reduces mapping error relative to satellite-only approaches
because it preserves structural anchors while providing full spatial coverage (Wang et al., 2023).

Temporal depth improves robustness and enables change detection: multi-epoch LiDAR or GEDI combined
with continuous optical and SAR time series captures biomass trajectories (decline and recovery) and stabilizes
estimates across variable acquisition conditions (Loh et al., 2022; Musthafa & Singh, 2022; Naik et al., 2021).
Recent reviews highlight multi-temporal, multi-platform fusion as a key route to accurate and stable long-term
monitoring (Balestra et al., 2024). More specifically, explicitly integrating seasonal and disturbance-driven
variability is critical for accurately quantifying biomass dynamics and carbon flux in Melaleuca wetlands. These
ecosystems are subject to strong seasonal hydrological pulses and periodic disturbances (e.g., fire or selective
logging), which significantly alter AGB over short time scales. By leveraging dense time-series data from
sensors like Sentinel or Landsat, it is possible to move beyond static AGB maps to dynamic monitoring systems.
Methodologies such as time-series change detection and break-point analysis can identify the timing and
magnitude of biomass loss or gain, providing crucial information for carbon accounting and management
interventions (DeVries et al., 2015; Zhu, 2017). Therefore, future modeling roadmaps should prioritize the
integration of these temporal metrics as predictive covariates to capture the full spectrum of AGB variability.

Practical note: choose sensor stacks according to scale and objective-use UAV/airborne LiDAR where detailed
structural inference and local allometry are required; use GEDI/L-band + Sentinel-1/2 with LiDAR strips for
regional mapping; and incorporate multi-temporal series when the goal is change detection or long-term
monitoring.

Model Applicability to Melaleuca forests

Applicability of remote-sensing AGB models in Melaleuca ecosystems depends critically on (i) the match
between training data and target domain, (ii) the sensor stack and degree of structural supervision, and (iii) the
inclusion of environmental covariates that capture peat/acid-sulfate dynamics. Models trained on mineral-soil
forests or on limited site conditions frequently underperform when transferred to peatland Melaleuca stands
because of systematic differences in allometry, soil dielectric properties, and hydrological regime (Huy et al.,
2016; Nam et al., 2016; Tran et al., 2015).

Scaffolded multi-sensor models (LiDAR/GEDI + L-band SAR + Sentinel-2) are most applicable when structural
anchors overlap the target domain and field plots adequately sample the principal substrate and hydrological
classes. Under these conditions, tree-based ensembles and well-regularized deep architectures typically yield
transferable estimates with explicit uncertainty quantification (Nguyen et al., 2024; Wang et al., 2023).

Key limitations and risk factors.

Domain shift: differences in peat depth, salinity/acid sulfate status, hydrological alteration (drainage canals) and
stand age/density create systematic biases if absent from training data.

Sensor saturation and structural heterogeneity: optical and C-band SAR indicators saturate at high biomass;
L-band and LiDAR mitigate but do not fully remove ambiguity in complex canopies (Zadbagher et al., 2024).

Sparse structural supervision: where LiDAR/GEDI coverage or field plot density is low, expect larger
extrapolation errors and spatially clustered uncertainty.

Recommended pre-deployment checks.

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025

www.rsisinternational.org Page 478

Domain diagnostic: compare distributions of key predictors (height metrics, hydroperiod proxies, peat depth,
spectral indices) between training and target areas; flag areas with large covariate shift.

Scaffold availability: ensure structural anchors (LiDAR strips, GEDI, representative plots) exist across main
substrate/hydrological strata; if absent, restrict inference or increase uncertainty.

Pilot external validation: reserve independent LiDAR strips or holdout regions for leave-one-region-out testing
to estimate realistic transfer error.

Mapping products and reporting.

Deliverables should include wall-to-wall AGB map plus (i) pixelwise (or grid) uncertainty estimates (e.g.,
prediction intervals or quantile maps), (ii) stratified error summaries by peat/soil class and hydrological class,
and (iii) a short “usage note” that identifies areas where models extrapolate beyond training support. Explicitly
report validation protocol (spatial/spatio-temporal blocking), sample sizes per strata, and any preprocessing
choices that materially affect inference (e.g., hydrologic digital elevation model [DEM] flattening).

Practical thresholding guidance.

- Treat regions with no structural scaffolding and with strong covariate shift as low confidence and avoid issuing
fine-scale AGB estimates without additional data collection.

- Use RF/GBM baselines for rapid operational mapping and reserve DL for contexts with dense LiDAR/plot
supervision. For formal reporting, always accompany maps with stratified uncertainty and a clear statement of
transferability limits.

Synthesis And Implications for Melaleuca Wetlands

Integrative overview

This integrative overview synthesizes the principal methodological insights from Sections 2.1 - 2.4 and
highlights their practical implications for mapping Melaleuca wetlands. Across studies, highest predictive
performance is achieved by scaffolded, multi-sensor approaches that combine structural information
(LiDAR/GEDI) with radar (notably L-band) and optical inputs, while rigorous spatial or spatio-temporal
validation is essential to avoid optimistic accuracy estimates (Nguyen et al., 2024; Roberts et al., 2017; Zhang et
al., 2019, 2020). Environmental covariates tied to peat and hydrological dynamics (peat depth, hydroperiod,
salinity, microtopography) consistently improve model transferability when they are represented in training data
(Huy et al., 2016; Tran et al., 2015; Zadbagher et al., 2024).

Key takeaways:

Scaffolded fusion is central: Use LiDAR/GEDI strips or footprints as structural anchors and extrapolate with
Sentinel-1/2 and L-band SAR for wall-to-wall mapping.

Validation defines realism: Spatial or spatio-temporal blocking (or leave-one-region-out tests) should be
standard for hyperparameter selection and performance reporting to estimate true transfer error.

Model choice should match data and compute: RF/XGBoost/LightGBM are robust baselines for most
operational contexts; deep learning is justified when dense structural supervision and large, diverse training
samples exist.

Environmental strata matter: Always stratify results (and report errors) by peat/soil/ hydrological classes to
expose heterogeneous performance and inform management use.

Report uncertainty and limits: Deliverables must include stratified uncertainty maps and concise usage notes that
identify low-confidence extrapolation zones.

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025

www.rsisinternational.org Page 479

For implementation guidance and worked examples that operationalize these principles, see Table 1, Table 2,
and Section 3.2.3.

Implications for Melaleuca (with hydrology - geomorphology - soil variables)

Synthesis and principal recommendations.

For Melaleuca-dominated wetlands, reliable AGB mapping requires workflows that (i) anchor remote-sensing
predictions in locally derived allometry (DBH - height - wood density), (ii) employ multi-sensor fusion with
emphasis on structural scaffolds (LiDAR or GEDI) and L-band SAR to mitigate optical/C-band saturation, (iii)
explicitly incorporate environmental covariates reflecting peat and acid-sulfate dynamics, and (iv) evaluate
transferability using spatial or spatio-temporal blocking across hydrological and management strata (Bui et al.,
2024; Huy et al., 2016; Luo et al., 2024; Musthafa & Singh, 2022; Nam et al., 2016; Tran et al., 2015).

Table 2. Per-study summary

No Reported metric(s) (best / representative) Sources

1 SSAE (deep model): R2 = 0.935, RMSE = 15.67 Mg·ha-1 Zhang et al. (2019)

2 Best performing (CatBoost/aggregated): R2 ≈ 0.72, RMSE = 45.63 Mg·ha-1
(CatBoost aggregated). Ensemble/tree-based mean R2 ≈ 0.69 - 0.71, RMSE ≈ 46 -

48 Mg·ha-1

Zhang et al. (2020)

3 Tent_ASO_BP (NN): R2 = 0.74, RMSE = 11.54 Mg·ha-1 (best configuration
reported). Comparators: RF R2 = 0.54 (RMSE 21.33), SVR R2 = 0.52 (RMSE

17.66), PLSR R² = 0.50 (RMSE 16.52)

Chen et al. (2023)

4 Reported model range R2 ≈ 0.615 - 0.754. Best RF: R2 = 0.754; reported MAE =
78.5 Mg·ha-1, %RMSE = 13.57% (abstract)

Nguyen et al.
(2024)

5 Best (SVM reported): R2 = 0.70, RMSE = 83.65 Mg·ha-1, MAE = 74.43 Mg·ha-1 -
highlights lower accuracies in structurally complex/high-biomass peat forests

Zadbagher et al.
(2024)

6 Combination (Sentinel-2A + ALOS-2 PALSAR-2), best model (SVR): R2 = 0.73,
RMSE = 38.68 Mg·ha-1 (SVR, Sentinel + ALOS)

Vafaei et al. (2018)

7 LiDAR-based (UAV strip) model (larch): R2 = 0.923, RMSE = 13.92 Mg·ha-1
(leave-one-out CV). Sentinel-based models (using LiDAR sampling) achieved

LiDAR-validation accuracies up to ~ 0.74 - 0.79 (R2 or % accuracy reported) and
Sentinel-based RMSEs (field vs LiDAR validation sets)

Wang et al. (2023)

Key environmental predictors and mechanistic role

Below are the predictor groups we recommend including as covariates or stratification layers; a detailed list with
measurement/derivation notes is provided in Table 3.

- Hydrology: inundation duration/hydroperiod, water-table depth, flood timing, and distance to canals/ditches (as
a proxy for drainage alteration). The hydrological regime influences canopy vitality and stem allometry, while
also modulating dielectric and optical signals through its effects on moisture content; omitting hydrological
metrics partly explains cross-site failures in transferability (Dang et al., 2022; Huy et al., 2016; Nguyen et al.,
2016).

- Soils and peat characteristics: peat depth, bulk density, soil salinity/acidity (acid sulfate indicators), and texture.
These substrate properties influence growth rates, wood density, and electromagnetic contrasts

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025

www.rsisinternational.org Page 480

(microwave/optical), and therefore strongly affect both predictive bias and generalization from mineral soils to
peatlands (Huy et al., 2016; Kappas, 2020; Tran et al., 2015).

- Geomorphology and microtopography: relative elevation (from hydrologically corrected DEMs), local
slope/curvature, and distance to levees/ridges. In very flat peat landscapes even small elevation differences can
control hydroperiod and vegetation structure; including these covariates materially improves spatial
transferability (Ngo et al., 2023; Nguyen and Nguyen, 2017).

Table 3: Public datasets for environmental covariates

Covariate
class

Variable
examples

Dataset Provider Access Primary
citation

Hydrology
(hydro-peri

od)

Water
occurrence,
seasonality,
recurrence

Global
Surface

Water (v1.4)

EC JRC https://global-surface-water.appsp
ot.com/download

Pekel et al.
(2016)

Hydrology
(networks)

Flow
accumulation/

direction,
distance‑to‑

channel

HydroSHED
S (core

products v1)

WWF/
USGS

consortium

https://www.hydrosheds.org/prod
ucts

HydroSHE
DS

Technical
Doc.

(2022)

Topograph
y (DEM)

Elevation,
slope, TWI

Copernicus
DEM

GLO‑30

ESA/
Copernicus

https://dataspace.copernicus.eu/ex
plore-data/data-collections/copern
icus-contributing-missions/collect

ions-description/COP-DEM

Soils
(texture/

peat)

Sand/silt/clay;
soil class;

proxies for peat

SoilGrids
250 m (v2.0)

ISRIC https://soilgrids.org/ Hengl et al.
(2017); de
Sousa et al.

(2021)

Wetlands/
peat extent

Tropical
wetlands and

peatland
likelihood

Tropical
wetlands/pea

t model

Gumbricht
et al.

- Gumbricht
et al.

(2017)

Coastal
wetland

(optional)

Mangrove
extent (blue‑

carbon context)

Global
Mangrove

Watch (v3.0)

JAXA/
Partners

https://www.globalmangrovewatc
h.org/

Bunting et
al. (2018)

Practical modelling roadmap

Scaffold and local allometry. Acquire or identify structural anchors (UAV/airborne LiDAR strips, GEDI
footprints) and derive local DBH - height - wood density relationships where possible to translate structure →
AGB.

Predictor fusion. Combine structural scaffolds with wall-to-wall Sentinel-1/2 and L-band SAR (when available)
plus the hydrology/peat/geomorphology layers listed above (Section 2.2, Table 1). Late feature fusion into
tree-based ensembles (RF/ XGBoost/LightGBM) makes a robust operational baseline; escalate to DL when
dense structural supervision and large training sets exist.

Validation and transfer testing. Use spatial and spatio-temporal blocking (block sizes guided by empirical

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025

www.rsisinternational.org Page 481

autocorrelation) and reserve independent LiDAR strips or holdout regions (leave-one-region-out) to quantify
realistic out-of-domain error (Roberts et al., 2017; Valavi et al., 2019). Stratify validation by peat/soil/hydro
class.

Uncertainty and reporting. Produce wall-to-wall AGB maps accompanied by pixel-wise uncertainty (quantiles
or prediction intervals), stratified error summaries, and a concise usage note identifying low-confidence
extrapolation areas (Musthafa & Singh, 2022; Ngo et al., 2023).

Operational cautions and decision rules

In regions lacking structural scaffolds and showing strong covariate shift (e.g., substantial differences in peat
depth or hydroperiod vs. training sites), treat fine-scale AGB estimates as low confidence and prioritize targeted
LiDAR/plot collection before operational mapping.

When computational or sample constraints exist, favor tree-based ensemble baselines (RF/GBM) and report
their limitations explicitly; report RMSE in Mg·ha-1 and include stratified error tables.

Always document preprocessing choices that affect hydrologic/geomorphic predictors (e.g., hydrologic DEM
flattening, peat depth interpolation methods), since such choices materially influence extrapolation behavior.

Explicitly integrating hydrology, geomorphology, and soil/peat variables into scaffolded, multi-sensor
modelling workflows is essential for producing transferable and actionable Melaleuca AGB products. For
implementation templates, code snippets, and recommended predictor derivations, see Section 3.2.3, Table 1,
and Table 3.

Temporal and disturbance factors

Seasonal dynamics and discrete disturbances (e.g., floods, fires, harvesting, storm damage) strongly influence
AGB patterns and carbon fluxes in Melaleuca wetlands. To account for these effects, we extend the framework
with time‑aware predictors and validation:

Multi‑temporal stacks. Build seasonal/monthly composites from Sentinel‑1 and Sentinel‑2 (e.g., pre‑flood,
peak‑flood, post‑flood) and include temporal statistics (median, IQR, trend) as features; demonstrations in the
Mekong Delta show the value of dense SAR/optical time series for flood hydrology (Lam et al., 2023; Tran et al.,
2022).

Hydrological regime dynamics. Derive flood frequency, duration, and timing from multi‑year water masks and
SWIR‑based moisture anomalies to capture inter‑annual variability; global surface‑water seasonality layers
provide a robust baseline (Pekel et al., 2016)..

Disturbance proxies. Integrate fire occurrence/burned area, logging footprints, and storm tracks; encode recency
(days since event), intensity, and cumulative disturbance history. Validated burned‑area products and algorithms
support time‑series disturbance mapping in tropical peatlands (Boschetti et al., 2019; Giglio et al., 2018).

Space–time validation. Complement spatial blocking with temporal or space-time blocked cross‑validation
(train on years t…t-k, test on t+1) to assess robustness under seasonal shifts and event shocks (Roberts et al.,
2017; Valavi et al., 2019).

Change‑aware features. For flux‑relevant analyses, include Δ‑features (year‑to‑year change in SAR/optical
indices) and report bias/variance separately for disturbed vs. non‑disturbed strata.

Uncertainty reporting. Map higher predictive uncertainty for periods immediately following major disturbances
or transitional hydrological phases, and follow AGB unit/uncertainty conventions when leveraging GEDI
products (Kellner et al., 2023; Dubayah et al, 2022).

This time‑aware extension improves reliability of biomass estimates under dynamic wetland conditions and

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025

www.rsisinternational.org Page 482

supports standardized AGB mapping across Southeast Asian wetlands facing similar seasonal and disturbance
regimes.

Operational feasibility, cost‑effectiveness, and regional generalizability

Operational uptake of multi‑sensor AGB mapping depends on cost‑effectiveness and implementation feasibility.
At national MRV scales, monitoring and transaction costs can erode the net benefits of result‑based payments if
system design is not cost‑sensitive (Köhl, Neupane, & Mundhenk, 2020). Practical pathways therefore prioritize
open and routinely updated sensors - Sentinel‑1/2 and L‑band SAR mosaics - combined with reproducible
processing on cloud platforms to reduce hardware and maintenance burdens (Gorelick et al., 2017; Shimada et
al., 2014; JAXA&EORC, 2022). Method guidance from REDD+ MRV frameworks emphasizes transparent
protocols, adequate sampling, and uncertainty management to balance precision with affordability (Böttcher et
al., 2009; Herold et al., 2011; GOFC‑GOLD, 2011). In data‑ and capacity‑limited contexts, a tiered modeling
strategy - starting with tree‑based ensembles and escalating to deep learning only when wall‑to‑wall inputs and
measurable accuracy gains are present - helps contain costs while meeting reporting requirements (Roberts et al.,
2017; Valavi et al., 2019).

Generalizability beyond Vietnam is supported by shared ecological and data conditions across Southeast Asian
wetlands. Peat‑dominated lowlands in Peninsular Malaysia, Sumatra, and Borneo show comparable
hydrological regimes and disturbance histories, with region‑wide declines in peat swamp forest cover since the
1990s that motivate standardized, repeatable mapping (Miettinen et al., 2016; Mishra et al., 2021). At broader
scales, tropical wetland/peat distributions and pan‑tropical biomass products provide consistent reference layers
for stratification and benchmarking (Gumbricht et al., 2017; Avitabile et al., 2016; Tootchi, Jost, & Ducharne,
2019). In practice, transfer is achieved by harmonizing environmental strata (peat depth, soil texture, flood
regime), applying blocked space-time validation, and leveraging regional time‑series demonstrations from the
Mekong Delta for flood‑driven variability (Lam et al., 2023; Tran et al., 2022). This combination of open data,
cost‑aware design, and explicit uncertainty reporting strengthens regional comparability and helps standardize
future AGB mapping across Southeast Asian wetlands.

CONCLUSION AND RECOMMENDATIONS

Conclusion

This review consolidates advances in multi‑sensor remote sensing and machine learning for AGB estimation in
Vietnam’s Melaleuca wetlands and proposes a standardized framework of terminology, indicators, and
environmental covariates tailored to tropical peatland ecosystems. By unifying key metrics such as canopy
structural parameters, hydro‑geomorphological indices, and model validation criteria (R², RMSE), the study
enhances methodological consistency and reproducibility across different research settings. Beyond serving as a
synthesis, the framework provides a replicable roadmap for multi‑sensor data integration - spanning data
acquisition, feature engineering, modeling, validation, and uncertainty quantification. This approach strengthens
national‑level biomass monitoring and carbon accounting and helps standardize future AGB mapping across
Southeast Asian wetlands where similar ecological and data constraints prevail.

Future Recommendations

We recommend using LiDAR/GEDI strips or footprints as calibration scaffolds, extrapolating wall-to-wall with
Sentinel-1/2 and L-band SAR, adopting late feature fusion with RF/XGBoost/LightGBM as robust baselines,
and escalating to compact deep learning only when dense structural supervision is available, and reporting
uncertainty (prediction intervals/quantiles) with stratified errors by peat/soil/hydrological classes.

To bridge the gap between technical advancement and real-world application in carbon management, future
efforts should focus on developing open-access tools and pre-trained models. We recommend leveraging
cloud-based geospatial platforms such as Google Earth Engine (GEE), which offers cost-effective and
operationally feasible Big Earth Data processing capabilities, particularly crucial for resource-constrained
governmental agencies (Gorelick et al., 2017). Specifically, developing a user-friendly workflow within GEE

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025

www.rsisinternational.org Page 483

can automate the complex pre-processing steps required for multi-sensor data fusion (e.g., Sentinel-1/2 and
GEDI). Furthermore, to promote transferability and reduce modeling time, the research community should
prioritize the public release of standardized training datasets and pre-trained tree-based machine learning models
(e.g., RF) via open-source repositories (e.g., GitHub/GEE Apps) (Wu, 2020). This approach will enable local
users to rapidly generate high-accuracy AGB maps, integrating crucial uncertainty parameters (Amitrano et al.,
2023), thereby directly supporting more transparent conservation planning and greenhouse gas inventory
reporting.

To establish the robustness and generalizability of remote sensing and machine learning methodologies, future
studies must extend beyond the localized scope of Vietnam’s Melaleuca wetlands. We recommend conducting
multi-regional comparative studies that evaluate the performance of AGB models calibrated in the Mekong
Delta against other regional tropical peat swamp forest ecosystems, such as those in Borneo (Indonesia) or the
Malay Peninsula (Malaysia). These ecosystems present similar vegetation structures and geochemical
conditions but often exhibit a higher range of AGB saturation, providing a necessary stress test for the algorithms
(Lohberger et al., 2013; Zadbagher et al., 2024). This comparison should specifically analyze how critical
environmental covariates - such as peat depth, seasonal hydrology, and salinity/acid-sulfate conditions -
influence model accuracy and bias across different regions. By quantifying these differences, researchers can
develop adaptive AGB models capable of self-adjusting based on region-specific input data, thereby maximizing
their utility for carbon accounting at a broader scale.

To transform AGB maps from a research tool into a reliable decision-support document, prioritizing the further
standardization of uncertainty quantification is essential. Future studies should move beyond merely reporting
aggregate statistics like RMSE and R2. We recommend adopting a comprehensive framework to assess AGB
map accuracy, including three essential elements (Weisbin et al., 2014; Sannier et al., 2022):

Prediction intervals (PIs) and quantiles: Providing a point estimate for AGB is insufficient. Studies must
calculate and map 95% prediction intervals (95% PIs) or other quantiles for every pixel. This transforms the map
from a single-value assertion into a statement of spatial confidence, transparently communicating the risk of
over- or underestimation, which is particularly crucial in areas with dense canopies where signal saturation is
common.

Bias analysis: Conditional bias is a prevalent issue in machine learning models where input data are unevenly
distributed. Advanced statistical techniques (such as model-assisted statistical regression) must be used to model
and adjust for map bias (Sannier et al., 2022). This analysis specifically quantifies whether the model
systematically over- or under-predicts in specific areas, such as high-biomass peat swamp forests, thereby
facilitating the creation of bias-adjusted maps suitable for carbon reporting standards.

Uncertainty source decomposition: Explicitly analyze and decompose the main sources of uncertainty,
including: field measurement errors, allometric model errors, and remote sensing errors. This decomposition
helps prioritize future efforts to reduce overall uncertainty most effectively.

REFERENCES

1. Amitrano D., Giacco G., Marrone S., Pascarella A.E., Rigiroli M., & Sansone C. (2023). Forest
aboveground biomass estimation using machine learning ensembles: Active learning strategies for model
transfer and field sampling reduction. Remote Sensing, 15(21), 5138. https://doi.org/ 10.3390/ rs1521
5138

2. Araza A., de Bruin S., Herold M., Quegan S., Labriere N., Rodriguez-Veiga P., Avitabile V., Santoro M.,
Mitchard E.T.A., Ryan C.M., Phillips O.L., Willcock S., Verbeeck H., Carreiras J., Hein L., Schelhaa
M.J., Pacheco-Pascagaza A.M., da Conceição Bispo P., Laurin G.V., Vieilledent G., Slik F., Wijaya A.,
Lewis S. L., Morel A., Liang J., Sukhdeo H., Schepaschenko D., Cavlovic J., Gilani H., Lucas R. (2022).
A comprehensive framework for assessing the accuracy and uncertainty of global above-ground biomass
maps. Remote Sensing of Environment, 272, 112917. https://doi.org/10.1016/j.rse.2022.112917

3. Avitabile V., Herold M., Heuvelink G.B.M., Lewis S.L., Phillips O.L., Asner G.P., Armston J., Ashton
P.S., Banin L., Bayol N., Berry N.J., Boeckx P., de Jong B.H.J., DeVries B., Girardin C.A.J., Kearsley E.,

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025

www.rsisinternational.org Page 484

Lindsell J.A., Lopez-Gonzalez G., Lucas R., Malhi Y., Morel A., Mitchard E.T. A., Nagy L., Qie L.,
Quinones M.J., Ryan C.M., Ferry S.J.W., Sunderland T., Laurin G.V., Gatti R.C., Valentini R., Verbeeck
H., Wijaya A., Willcock S. (2015). An integrated pan‑tropical biomass map using multiple reference
datasets. Global Change Biology, 22(4), 1406–1420. https://doi.org/10.1111/gcb.13139

4. Balestra M., Marselis S., Sankey T.T., Cabo C., Liang X., Mokroš M., Peng X., Singh A., Stereńczak K.,
Vega C., Vincent G., & Hollaus M. (2024). LiDAR data fusion to improve forest attribute estimates: A
review. Current Forestry Reports, 10, 281-297. https://doi.org/10.1007/s40725-024-00223-7

5. Boschetti L., Roy D.P., Giglio L., Huang H., Zubkova M., & Humber M.L. (2019). Global validation of
the Collection 6 MODIS burned area product. Remote Sensing of Environment, 235, 111490.
https://doi.org/10.1016/j.rse.2019.111490

6. Böttcher H., Eisbrenner K., Fritz S., Kindermann G., Kraxner F., McCallum I., & Obersteiner M. (2009).
An assessment of monitoring requirements and costs of ‘Reduced Emissions from Deforestation and
Degradation’. Carbon Balance and Management, 4, 7. https://doi.org/10.1186/ 1750 -0680-4-7

7. Bui Q., Pham Q., Pham V., Tran V., Nguyen D., Nguyen Q., Nguyen H.D., Do N.T., & Vu V. (2024).
Hybrid machine learning models for aboveground biomass estimations. Ecological Informatics, 79:
102421. https://doi.org/10.1016/j.ecoinf.2023.102421

8. Bunting P., Rosenqvist A., Lucas R.M., Rebelo L.M., Hilarides L., Thomas N., Hardy A., Itoh T.,
Shimada M., & Finlayson C.M. (2018). The global mangrove watch - A new 2010 global baseline of
mangrove extent. Remote Sensing, 10(10), 1669. https://doi.org/10.3390/rs10101669

9. Chen Z., Du J., He Y., Liu Y., Zhang Z., Sun Q., & Sun S. (2023). Aboveground forest biomass
estimation using tent mapping atom search optimized backpropagation neural network with Landsat 8
and Sentinel-1A data. Remote Sensing, 15(24), 5653. https://doi.org/10.3390/rs15245653

10. Dang H.N., Ba D.D., Trung D.N., & Viet H.N.H. (2022). A novel method for estimating biomass and
carbon sequestration in tropical rainforest areas based on remote sensing imagery: A case study in the
Kon Ha Nung plateau, Vietnam. Sustainability, 14 (24), 16857-16857. https://doi.org/10.3390/ su1
42416857

11. DeVries B., Pratihast A.K, Verbesselt J., Kooistra L., & Herold M. (2015). Characterizing forest change
using community-based monitoring and Landsat time series. PLoS ONE, 11(3): e0147121. https://
doi.org/10.1371/journal.pone.0147121

12. Dubayah R.O., Armston J., Kellner J.R., Duncanson L., Healey S.P., Patterson P.L., Hancock S., Tang H.,
Bruening J.M., Hofton M.A., Blair J.B., & Luthcke S.B. (2022). GEDI L4A footprint level aboveground
biomass density, Version 2.1. ORNL DAAC, Oak Ridge, Tennessee, USA. https://
doi.org/10.3334/ORNLDAAC/2056

13. Englhart S., Franke J., Keuck V., & Siegert F. (2012). Aboveground biomass estimation of tropical peat
swamp forests using SAR and optical data. 2012 IEEE International Geoscience and Remote Sensing
Symposium, Munich, Germany. 6577-6580, https://doi.org/10.1109/IGARSS.2012.6352092

14. Giglio L., Boschetti L., Roy D.P., Humber M.L., & Justice C.O. (2018). The Collection 6 MODIS burned
area mapping algorithm and product. Remote Sensing of Environment, 217, 72–85. https://
doi.org/10.1016/j.rse.2018.08.005

15. GOFC‑GOLD (2011). A sourcebook of methods and procedures for monitoring and reporting
anthropogenic greenhouse gas emissions and removals caused by deforestation, gains and losses of
carbon stocks in forests remaining forests, and forestation (COP17‑1). GOFC‑GOLD Project Office,
Natural Resources Canada. https://redd.unfccc.int/uploads/63_33_redd_20120509_gofc-gold.pdf

16. Gorelick N., Hancher M., Dixon M., Ilyushchenko S., Thau D. & Moore R. (2017). Google Earth
Engine: Planetary-scale geospatial analysis for everyone. Remote Sensing of Environment, 202, 18–27.
https://doi.org/10.1016/j.rse.2017.06.031

17. Gumbricht T., Roman‑Cuesta R.M., Verchot L., Herold M., Wittmann F., Householder E., Herold N., &
Murdiyarso D. (2017). An expert system model for mapping tropical wetlands and peatlands reveals
South America as the largest contributor. Global Change Biology, 23(9), 3581–3599.
https://doi.org/10.1111/gcb.13689

18. Hengl T., de Jesus J.M., Heuvelink G.B.M., Gonzalez M.R., Kilibarda M., Blagotić A., Shangguan W.,
Wright M.N., Geng X., Bauer‑Marschallinger B., Guevara M.A., Vargas R., MacMillan R.A., Batjes
N.H., Leenaars J.G.B., Ribeiro E., Wheeler I, Mantel S., & Kempen B. (2017). SoilGrids250m: Global
gridded soil information based on machine learning. PLOS ONE, 12(2), e0169748. https:// doi.org/

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025

www.rsisinternational.org Page 485

10.1371/journal.pone.0169748
19. Herold M., Román‑Cuesta R.M., Mollicone D., Hirata Y., van Laake P., Asner G.P., Souza C., Skutsch

M., Avitabile V., & MacDicken K. (2011). Options for monitoring and estimating historical carbon
emissions from forest degradation in the context of REDD+. Carbon Balance and Management, 6, 13.
https://doi.org/10.1186/1750-0680-6-13

20. Huy B., Poudel K.P., Kralicek K., Hung N.D., Khoa P.V., Phuong V.T., & Temesgen H. (2016).
Allometric equations for estimating tree aboveground biomass in tropical dipterocarp forests of
Vietnam. Forests, 7(8), 180. https://doi.org/10.3390/F7080180

21. JAXA & EORC - Japan Aerospace Exploration Agency, and Earth Observation Research Center (2022).
Global 25 m Resolution PALSAR‑2 Forest/Non‑Forest Map (FNF) (Ver. 2.0.0) - Dataset Description.
https://www.eorc.jaxa.jp/ALOS/en/dataset/pdf/DatasetDescription_PALSAR2_FNF_v200a.pdf

22. Kappas M. (2020). Estimating the aboveground biomass of an evergreen broadleaf forest in Xuan Lien
nature reserve, Thanh Hoa, Vietnam, using SPOT-6 data and the random forest algorithm. International
Journal of Forestry Research, 1, 4216160. https://doi.org/10.1155/2020/4216160

23. Kellner J.R., Armston J., & Duncanson L. (2023). Algorithm theoretical basis document for GEDI
footprint aboveground biomass density. Earth and Space Science, 10(4), e2022EA002516.
https://doi.org/10.1029/2022EA002516

24. Khan M.N., Tan Y., Gul A.A., Abbas S., & Wang J. (2024). Forest aboveground biomass estimation and
inventory: Evaluating remote sensing-based approaches. Forests, 15(6), 1055. https://doi.org/ 10.3390/
f15061055

25. Köhl M., Neupane P.R., & Mundhenk P. (2020). REDD+ measurement, reporting and verification – A
cost trap? Implications for financing REDD+ MRV costs by result‑based payments. Ecological
Economics, 168, 106513. https://doi.org/10.1016/j.ecolecon.2019.106513

26. Lam C.N., Niculescu S., & Bengoufa S. (2023). Monitoring and mapping floods and floodable areas in
the Mekong Delta (Vietnam) using time-series Sentinel-1 images, convolutional neural network,
multi-layer perceptron, and random forest. Remote Sensing, 15(8), 2001. https://doi.org/ 10.3390/
rs15082001

27. Lehner B. (2022). HydroSHEDS Technical Documentation – Data Version 1.1. https://
data.hydrosheds.org/file/technical-documentation/HydroSHEDS_TechDoc_v1_4.pdf

28. Loh H.Y., James D., Ioki K., Wong W.V.C., Tsuyuki S., & Phua M.H. (2022). Estimating aboveground
biomass changes in a human-modified tropical montane forest of Borneo using multi-temporal airborne
LiDAR data. Remote Sensing Applications: Society and Environment, 28, 100821. https://
doi.org/10.1016/j.rsase.2022.100821

29. Luo M., Anees S.A., Huang Q., Xin Q., Qin Z., Fan J., Han G., Zhang L., & Shafri H.Z.M. (2024).
Improving forest above-ground biomass estimation by integrating individual machine learning
models. Forests, 15 (6), 975. https://doi.org/10.3390/f15060975

30. Miettinen J., Shi C., & Liew S.C. (2016). Land cover distribution in the peatlands of Peninsular Malaysia,
Sumatra and Borneo in 2015 with changes since 1990. Global Ecology and Conservation, 6, 67–78.
https://doi.org/10.1016/j.gecco.2016.02.004

31. Mishra S., Page S.E., Cobb A.R., Lee J.S.H., Jovani-Sancho A.J., Sjögersten S., Jaya A., Aswandi,
Wardle D.A. (2021). Degradation of Southeast Asian tropical peatlands and integrated strategies for their
better management and restoration. Journal of Applied Ecology, 58(7), 1370–1387.
https://doi.org/10.1111/1365-2664.13905

32. Musthafa M., & Singh G. (2022). Improving forest above-ground biomass retrieval using multi-sensor
L- and C-band SAR data and multi-temporal spaceborne LiDAR data. Forest Disturbance, 5, 822704.
https://doi.org/10.3389/ffgc.2022.822704

33. Naik P., Dalponte M., & Bruzzone L. (2021). Prediction of forest aboveground biomass using
multitemporal multispectral remote sensing data. Remote Sensing, 13(7), 1282. https://doi.org/
10.3390/rs13071282

34. Nam V.T., Kuijk M.V., & Anten N.P.R. (2016). Allometric equations for aboveground and belowground
biomass estimations in an evergreen forest in Vietnam. Research Article, 11(6), e0156827.
https://doi.org/10.1371/journal.pone.0156827

35. Ngo D.T., Nguyen H.D., Nguyen K.D.T., Dang C.K., Dang N.K., & Pham T.V. (2023). Application of
multispectral UAV to estimate mangrove biomass in Vietnam: A case study in Dong Rui commune,

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025

www.rsisinternational.org Page 486

Quang Ninh province. One Ecosystem, 8, e103760. https://doi.org/10.3897/oneeco.8.e103760
36. Nguyen H.K.L., & Nguyen B.N. (2017). Mapping biomass and carbon stock of forest by remote sensing

and GIS technology at Bach Ma National park, Thua Thien Hue province. Journal of Vietnamese
Environment, 8(2), 80-87. https://doi.org/10.13141/JVE.VOL8.NO2.PP80-87

37. Nguyen L.V., Tateishi R., Kondoh A., Sharma R.C., Nguyen H.T., To T.T., & Minh D.H.T. (2016).
Mapping tropical forest biomass by combining ALOS-2, LANDSAT 8, and field plots data. Land, 5(4).
https://doi.org/10.3390/LAND5040031

38. Nguyen T.P., Nguyen P.K., Nguyen H.N., Tran T.D., Pham G.T., Le T.H., Le D.H., Nguyen T.H., &
Nguyen, V. B. (2024). Evaluation of statistical and machine learning models using satellite data to
estimate aboveground biomass: A study in Vietnam tropical forests. Forest Science and Technology,
20(4), 370-382. https://doi.org/10.1080/21580103.2024.2409211

39. Pekel J.F., Cottam A., Gorelick N., & Belward A.S. (2016). High-resolution mapping of global surface
water and its long-term changes. Nature, 540, 418–422. https://doi.org/10.1038/nature20584

40. Roberts D.R., Bahn V., Ciuti S., Boyce M.S., Elith J., Guillera-Arroita G., Hauenstein S., Lahoz-Monfort
J.J., Schröder B., Thuiller W., Warton D.I., Wintle B.A., Hartig F., & Dormann C.F. (2017).
Cross-validation strategies for data with temporal, spatial, hierarchical, or phylogenetic structure.
Ecography, 40(8), 913–929. https://doi.org/10.1111/ecog.02881

41. Shimada M., Itoh T., Motooka T., Watanabe M., Shiraishi T., Thapa R., & Lucas R. (2014). New global
forest/non‑forest maps from ALOS PALSAR data (2007–2010). Remote Sensing of Environment, 155,
13–31. https://doi.org/10.1016/j.rse.2014.04.014

42. Tootchi A., Jost A., & Ducharne A. (2019). Multi‑source global wetland maps combining surface water
imagery and groundwater constraints. Earth System Science Data, 11(1), 189–220. https://doi.org/
10.5194/essd-11-189-2019

43. Tran D.B., Hoang T.V., & Dargusch P. (2015). An assessment of the carbon stocks and sodicity tolerance
of disturbed Melaleuca forests in Southern Vietnam. Carbon Balance and Management, 10, 15. https://
doi.org/10.1186/s13021-015-0025-6

44. Tran K.H., Menenti M., & Jia L. (2022). Surface water mapping and flood monitoring in the Mekong
Delta using Sentinel-1 SAR time series and Otsu threshold. Remote sensing, 14(22), 5721.
https://doi.org/10.3390/rs14225721

45. Vafaei S., Soosani J., Adeli K., Fadaei H., Naghavi H., Pham T.D., & Bui D.T. (2018). Improving
accuracy estimation of forest above-ground biomass based on incorporation of ALOS-2 PALSAR-2 and
Sentinel-2A imagery and machine learning: A case study of the Hyrcanian forest area (Iran). Remote
Sens., 10(2), 172. https://doi.org/10.3390/rs10020172

46. Valavi R., Elith J., Lahoz-Monfort J.J., & Guillera-Arroita G. (2019). BlockCV: An R package for
generating spatially or environmentally separated folds for k-fold cross-validation of species distribution
models. Methods in Ecology and Evolution, 10(2), 225-232. https://doi.org/10.1111/2041-210X.13107

47. Wang Y., Jia X., Chai G., Lei L., & Zhang X. (2023). Improved estimation of aboveground biomass of
regional coniferous forests integrating UAV-LiDAR strip data, Sentinel-1 and Sentinel-2 imageries. Plant
Methods 19, 65. https://doi.org/10.1186/s13007-023-01043-9

48. Weisbin C.R., Lincoln W., & Saatchi S. (2013). A systems engineering approach to estimating
uncertainty in above‐ground biomass (AGB) derived from remote‐sensing data. Systems Engineering,
17(3), 361–373. https://doi.org/10.1002/sys.21275

49. Wu Q. (2020). geemap: A Python package for interactive mapping with Google Earth Engine. Journal of
Open Source Software, 5(51), 2305. https://doi.org/10.21105/joss.02305

50. Yan Y., Lei J., & Huang Y. (2024). Forest aboveground biomass estimation based on unmanned aerial
vehicle-light detection and ranging and machine learning. Sensors, 24(21), 7071. https://doi.org/
10.3390/s24217071.

51. Zadbagher E., Marangoz M.A., & Becek K. (2024). Estimation of above-ground biomass using machine
learning approaches with InSAR and LiDAR data in tropical peat swamp forest of Brunei Darussalam.
iForest - Biogeosciences and Forestry, 17(3), 172-179. https://doi.org/10.3832/ifor4434-017

52. Zhang L., Shao Z., Liu J., & Cheng Q. (2019). Deep Learning Based Retrieval of Forest Aboveground
Biomass from Combined LiDAR and Landsat 8 Data. Remote Sensing, 11(12), 1459. https://doi.org/
10.3390/rs11121459

53. Zhang Y., Ma J., Liang S., Li X., & Li M. (2020). An evaluation of eight machine learning regression