01051000046 application of geospatial techniques in the assessment of tropical forest biomass reserves with conservation value in bac kan province
INTRODUCTION
Background of the research
Forest ecosystems act as both sources and sinks of atmospheric carbon dioxide (CO₂) and are essential for carbon sequestration and regulating the global carbon cycle (Yadav et al., 2022) Carbon reservoirs at local, regional, and national scales are critical for assessing the potential of various carbon sequestration strategies, thereby contributing to the reduction of atmospheric CO₂ accumulation and the prevention of global warming (Jat et al., 2022) Article 2.1 of the Kyoto Protocol articulates the commitments of participating nations to mitigate the impacts of climate change through the protection of greenhouse gas reservoirs (Pfaff et al., 2000) These measures include afforestation, reforestation, and the promotion of sustainable forest management, all aimed at optimizing the capacity for carbon uptake and sequestration and thus reducing atmospheric CO₂ concentrations (Skrzypczak et al., 2025) Currently, forests worldwide are estimated to store more than 660 billion tonnes of carbon, with an average of about 163 tonnes of carbon per hectare, equivalent to nearly 557 billion cubic meters of wood (FAO, 2022) Forests are integral to global climate regulation owing to their capacity to absorb and sequester carbon Tropical forests, in particular, are often described as the “lungs of the Earth”, providing essential ecosystem services Through photosynthetic gas exchange, they assimilate CO₂ from the atmosphere and release O₂, thereby contributing to the maintenance of global ecosystem balance Tropical ecosystems cover approximately 33 percent of the Earth’s terrestrial surface and play a pivotal role in the formation and stabilization of carbon reservoirs (Yuan et al., 2024) Tropical forests alone account for approximately 45 percent of the world’s total forest area, equivalent to 1.83 billion hectares out of 4.06 billion hectares, and contribute substantially to the uptake of carbon dioxide, sequestering an estimated 56 percent of atmospheric CO₂ (1.4 billion tonnes) (Srivastav, 2024) Beyond their climate regulation function, tropical forests provide a wide range of essential ecosystem services such as biomass production, carbon storage and sequestration, and they influence climatic conditions at both global and regional scales Accurate understanding and quantification of tropical forest biomass stocks are therefore critical for assessing their role in the worldwide carbon cycle
Ecosystems within tropical forests provide essential goods and services at scales from local to global They supply timber, fuelwood, and other forest products, regulate soil and water resources, sustain biodiversity, deliver atmospheric functions, and promote overall ecological health and ecotourism Nevertheless, efforts to mitigate rising atmospheric CO₂ concentrations have led to conservation initiatives that focus exclusively on the carbon storage of these ecosystems (Wang et al., 2025)
To achieve success, whether in the voluntary carbon market or under climate change agreements following the Kyoto Protocol, within frameworks aimed at reducing emissions from deforestation and forest degradation, as well as promoting sustainable forest management, conservation, and the enhancement of forest carbon stocks (REDD+), it is essential to use a global, low-cost, and reliable approach to measure and monitor accumulated carbon stocks across large regions (Ahmed et al., 2025) However, accurately characterizing the structural attributes of these inaccessible ecosystems through conventional field inventory methods is exceedingly difficult because tropical forests host a vast array of species and exhibit structural complexity that varies across spatial and temporal scales (Bohlman, 2024) Field-based inventories of carbon stocks across large tropical forest areas face logistical challenges, are time-consuming, and are hard to replicate spatially for monitoring growth Existing long-term and short-term field plots in tropical forests provide valuable information on forest function and attributes, but they are too few to fully represent the dynamics of tropical forest ecosystems and carbon stocks
Spaceborne and airborne remote sensing technologies and sensor platforms are evolving rapidly and are increasingly used to gather information quickly, reliably, and consistently over large areas, thus supplementing infrequent and geographically limited field inventories (Bohlman, 2024) Understand the functions, structure, and productivity of tropical forest ecosystems, and document changes in forest attributes such as leaf area structure, basal area, stem volume, basal area-weighted mean diameter, and basal area-weighted mean height across different spatial and temporal scales (Fahlvik and Bửhlenius, 2025; Palace et al., 2015) Forest aboveground biomass (AGB) estimates come from various methods and sources, including field measurements, national forest inventories, administrative statistical data, model outputs, and regional satellite imagery products
Remote sensing cannot directly measure biomass; however, it is essential for accurately assessing and updating forest structure to estimate carbon stocks and fluxes (Quegan et al., 2019) Field-based inventory methods for estimating aboveground biomass are time-consuming, have low sampling density, lack spatial consistency, and often do not provide accurate AGB estimates in forest ecosystems (Van Winckel et al., 2025) Nevertheless, these methods remain essential and have played a key role in developing allometric equations for estimating biomass and carbon stocks, especially when combined with remote sensing technologies Remote sensing can surpass the limitations of traditional AGB measurement methods, particularly in hard-to-reach tropical forests, by providing cost-effective, repeatable, spatially continuous, and precise observations Currently, remote sensing-based methods are moving toward independent use and need dedicated AGB modeling frameworks Therefore, traditional measurement approaches, such as vegetation biophysical attributes, are being combined with multiple remote sensing datasets to measure and track aboveground biomass (AGB) stocks across various spatial and temporal scales
Assessing the cumulative carbon stock (CS) is essential for understanding the amount of carbon in biomass and the potential of forest ecosystems to serve as a carbon sink (Xu et al., 2025) Vietnam’s accession to the Kyoto Protocol highlights its commitment to global climate change mitigation and adaptation, even though, as a non-Annex I developing country, it is not required to undertake direct emission reductions Vietnam has taken proactive steps to address climate change impacts, such as the national forest inventory and assessment program This program, carried out every five years using traditional survey methods, aims to update and accurately assess forest area, forest quality, and the carbon sequestration capacity of forest ecosystems However, forest inventories in Vietnam’s Northeast region encounter significant challenges due to site heterogeneity and restricted accessibility, particularly in mountainous areas and primary forests In particular, Bac Kan province, a mountainous inland area in the center of the Northeast region of Vietnam, exhibited the highest national forest cover rate of 73.38 percent in 2023 (Thi Nhung et al., 2024) Bac Kan is home to diverse tropical forest ecosystems rich in natural resources, playing a vital role in air regulation and mitigating the impacts of climate change Therefore, this thesis aims to provide comprehensive information on aboveground biomass (AGB) estimates and cumulative carbon stock (CS) across forest ecosystems with varying conservation values.
Research rationale
• Scientific rationale: The findings of this dissertation significantly advance the efficiency of geospatial data processing and the analysis of forest attributes, including structural, morphological, and socio-economic aspects This supports the quantification of aboveground biomass (AGB) stocks and the carbon sequestration potential of tropical forests The research introduces a non-invasive method for estimating AGB stocks and cumulative carbon stock (CS) in the tropical forest ecosystems of Bac Kan Province, Vietnam, furthering the goals of sustainable forest management conservation
• Practical basis: This study emphasizes the critical role of quantifying biomass and carbon storage capacity in tropical forest ecosystems to support participation in the carbon credit market The research specifically focuses on selecting and integrating indices from multi-source remote sensing data, terrain information, and field measurements to noninvasively estimate aboveground biomass (AGB) and cumulative carbon stock (CS) in tropical forests The findings offer a scientific basis to assist forest managers in spatial organization and planning adjustments Additionally, they align with the development strategy of the newly consolidated province, effectively promoting both voluntary and compliance carbon initiatives markets.
Research question
• How can multi‐source remote sensing data, topographic information, and field measurements be integrated into a Cubist model to estimate aboveground biomass (AGB) and cumulative carbon stock (CS) in tropical forests?
• Which remote sensing indices have the most significant impact on the accuracy of the Cubist model when estimating AGB and CS in Bac Kan Province?
• How can the study’s findings help improve forest management policies and encourage participation in the global carbon credit market?
Research aims and objectives
• Objectives: Research integrating multi-source remote sensing data, topographic information, and field measurements to estimate aboveground biomass (AGB) and cumulative carbon stock (CS) in tropical forests, thereby supporting forest conservation, monitoring, and restoration efforts, as well as the development of sustainable forestry policies and participation in the global carbon credit markets
• Research aims: To achieve the research objectives, the thesis has conducted the following primary research tasks:
✓ Establish a scientific foundation for using geospatial technologies to non- invasively estimate aboveground biomass (AGB) and cumulative carbon stock (CS) in tropical forest ecosystems;
✓ Investigate and choose a set of multi-source remote sensing indices to develop models for estimating aboveground biomass (AGB) and cumulative carbon stock (CS);
✓ Use geospatial technologies to develop an aboveground biomass (AGB) estimation model and generate spatial distribution maps of AGB and cumulative carbon stock (CS) for Bac Kan province;
✓ Analysis and assessment of the cumulative carbon stock in high-value tropical forest ecosystems in Bac Kan province during the period 2016-2025;
✓ Propose strategies for sustainable forest protection, management, and development to encourage participation in the global carbon credit market, especially considering the expanded development scope after the merger of Bac Kan and Thai Nguyen provinces.
Scope of this research
• Spatial scope : The thesis conducts empirical research in Bac Kan Province (Vietnam), which covers an approximate natural area of 4,860 km² Bac Kan is a mountainous province located in the central part of the country’s Northeast region The study area spans from 12°42′36″ to 13°41′28″ North latitude and from 108°40′40″ to 109°27′47″ East longitude
• Temporal scope : The thesis used geospatial data to estimate the spatial distribution of aboveground biomass (AGB) and cumulative carbon stock (CS) in the tropical forest ecosystems of Bac Kan province from 2016 to 2025
• Scientific scope : The primary research method of this thesis involves developing models to estimate aboveground biomass (AGB) and cumulative carbon stock (CS) These models are based on analyzing indices derived from multi-source remote sensing data, topographic datasets, and field samples collected using geospatial methods technologies.
Significance of Thesis
• Scientific significance : This thesis advances geospatial data processing methods by effectively integrating remote sensing technologies with machine learning algorithms, thereby strengthening the scientific basis of the research These combined tools enable truly comprehensive analyses, merging spectral, structural, and temporal data to generate detailed maps of aboveground biomass, cumulative carbon stock, and land-use dynamics Such insights directly support sustainable development strategies, underpin carbon accounting methods, guide REDD+ planning, and shape conservation policies Additionally, by providing dynamic scenario simulations and long-term projections, the framework supplies decision- makers with data-driven pathways for optimized land-use zoning, ecosystem restoration, and resource allocation, reinforcing the scientific foundation for resilient spatial management planning
• Practical significance : The experimental results of this thesis not only support the underlying scientific theories but also improve understanding of how geospatial technologies can be used to map the spatial distribution of aboveground biomass (AGB) and cumulative carbon stock (CS) Additionally, the study offers practical value through an interdisciplinary approach, combining science and technology, environmental planning, and policy management to enhance the AGB and CS estimation processes in support of sustainable development.
Structure of Thesis
The thesis is organized into five chapters, each covering a key aspect of the research Chapter 1 establishes the foundation by presenting the study’s context, explaining its theoretical basis, formulating the main research questions, setting objectives and content, defining the scope, and outlining the overall structure of the dissertation
Chapter 2 provides a thorough review of literature on aboveground biomass (AGB) and cumulative carbon stock (CS) in tropical forests, covering core concepts, the ecological roles and ecosystem services of these forests, established and emerging methods for biomass and carbon evaluation, techniques for measuring tree biophysical parameters, and the use of remote sensing technologies
Chapter 3 describes the research methodology, including the physical and ecological features of the Bac Kan Province case study area It covers the identification and preprocessing of data sources such as multi-source satellite imagery, topographic datasets, and field measurements The chapter also explains the analytical framework, especially the development and validation of the Cubist modeling approach
Chapter 4 presents and analyzes the experimental findings, reporting on model performance metrics, mapping the spatial distribution of AGB and CS, evaluating key remote sensing variables, and discussing the implications of these results for forest monitoring and management
Finally, Chapter 5 summarizes the main conclusions, emphasizes the study’s contributions to sustainable forestry policy and the carbon-credit market, provides specific recommendations for long-term forest protection and spatial planning, and suggests directions for future research.
LITERATURE REVIEW
Tropical forests and their ecosystem services
Forests worldwide are considered habitats that support biodiversity, playing a key role in maintaining ecological balance Among these, tropical forests hold a critical position and consistently attract global attention (Saha, 2019) They provide habitat for most of Earth’s plant and animal species and form highly complex networks of ecological interactions, both among species and between organisms and their environment Tropical forests exhibit exceptional diversity, with tree species counts often reaching about 300 species per hectare These forests mainly occur in Southeast Asia, the Amazon Basin, the Congo Basin, as well as in western Amazonia and northern Borneo They are highly diverse and complex ecosystems, spanning from lowland areas to high mountains, with species distribution and forest structure changing based on climate and geographic location
Notably, tropical forests exhibit vibrant biodiversity and encompass a variety of ecosystem types such as mangrove forests, moist tropical forests, tropical rainforests, and semi-deciduous (seasonal) forests These ecosystems play a crucial role in biodiversity conservation, as the majority of associated biological resources are found and flourish within tropical forest environments (Armenteras et al., 2015; Soto-Navarro et al., 2020) Tropical forests provide a wide range of ecosystem services that are essential for climate regulation at local, regional, and global levels They help control temperature and humidity, reduce the impact of extreme weather events, manage the water cycle, protect watersheds and vegetation, support streamflow, and prevent soil erosion Additionally, tropical forests contain an extensive reservoir of genetic resources that remains unexplored, primarily offering significant potential for natural product discovery and biodiversity research Importantly, these forests are key to biomass production, carbon storage, and sequestration, thereby playing an essential role in regulating atmospheric greenhouse gas concentrations
Figure 2.1 Illustration of the global spatial distribution of forest ecosystems (FAO)
A thorough understanding of forests requires examining the values they provide to humanity, which are affected by ongoing changes in tropical forest ecosystems, including natural disturbances, deforestation, and land degradation (Buřivalová et al., 2023; Moges et al., 2024) As forest cover decreases, biodiversity loss speeds up over time, especially in “hotspots” with rapid land-use change Models indicate that fully protecting and restoring all protected areas could prevent 18% of species from going extinct Conversely, limiting conservation to current protected areas for ten years might result in the loss of up to 40% of species globally (Schmitz et al., 2023) Tropical deforestation not only depletes natural resources and harms ecosystem services but also releases large amounts of CO₂ and other greenhouse gases, thereby accelerating global climate change (Harris et al., 2021) Therefore, protecting forests within national parks and similar protected areas has become a top priority for international organizations, government agencies, and local communities Amid increasing climate challenges, current conservation efforts focus on boosting the carbon-storage capacity and reducing carbon emissions of tropical forests.
Tropical forest carbon storage and biomass assessment
2.2.1 Forest carbon storage and sequestration
Forests play a pivotal role in the global carbon cycle, as they store the largest terrestrial carbon sinks and continuously mediate carbon exchange between the Earth’s biosphere and atmosphere (Nzabarinda et al., 2025) Through photosynthesis, trees absorb atmospheric CO₂ and convert it into carbon-based organic compounds that are stored in biomass, including leaves, flowers, stems, and roots Meanwhile, through autotrophic respiration, trees use O₂ to break down some of these compounds and release CO₂ back into the atmosphere Over time, individual trees and entire forest stands store large amounts of carbon in their biomass When trees die, part of this stored carbon is released into the soil, where it can remain sequestered for thousands of years, while the rest is emitted into the atmosphere (Tyagi et al., 2025) Carbon release can happen quickly through wildfires or gradually as leaves, branches, and other organic materials break down Additionally, some woody biomass may continue to store carbon for a long time, either because it decomposes slowly or because it is harvested and turned into construction or industrial products The carbon in these end-use wood products will eventually be released, but when it happens varies greatly depending on the type and lifespan of the product
The processes of vegetation growth, mortality, and decomposition control carbon stocks in forests When total carbon emissions to the atmosphere exceed uptake, a forest becomes a net CO₂ source; conversely, when uptake exceeds emissions, it acts as a CO₂ sink Whether a specific forest is classified as a source or a sink depends on the spatial and temporal scales of assessment Globally, forests are considered carbon sinks, although the amount stored varies by region and geographic context (Peng et al., 2025) According to the United States Environmental Protection Agency (EPA), forest carbon stocks are categorized into seven primary sinks: (i) aboveground biomass, which includes stems, branches, foliage, bark, and propagules; (ii) belowground biomass, consisting of roots with diameters greater than 2.0 mm; (iii) deadwood, composed of standing and fallen woody debris; (iv) litter, which includes leaves and acceptable woody debris (diameter less than 7.5 cm) on the forest floor; (v) soil organic carbon up to a depth of 1 meter, divided into mineral and organic fractions; (vi) harvested wood products still in use; and (vii) harvested wood products removed from the forest ecosystem
Figure 2.2 Schematic diagram of forest carbon stock cycling and carbon transfers
Carbon residence times in forest pools are influenced by climatic conditions, hydrological regimes, nutrient availability, stand age, and ecosystem type (Button et al., 2022) These drivers influence the rates of carbon accumulation and release, causing significant variability in carbon stocks across different forest ecosystems As a result, carbon storage patterns can vary significantly between forest biomes Globally, boreal forests (taiga) cover about 29% of the world’s forest area and hold nearly one-third of the total biomass carbon In comparison, temperate forests cover a smaller land area and store less carbon Meanwhile, tropical forests make up almost 50% of the global forest area but contain over half of the world’s forest biomass carbon In tropical biomes, carbon is relatively evenly distributed between aboveground and belowground biomass, with a larger proportion stored above ground Although they make up only about 30% of the Earth’s forest cover, tropical forests hold up to 50% of the world’s total biomass carbon, highlighting their vital role in climate regulation The total carbon stocks in tropical forests across South America, sub-Saharan Africa, Southeast Asia, and Oceania reach 247 billion tonnes, with 193 billion tonnes in aboveground biomass (AGB) and 54 billion tonnes in belowground biomass (BGB) (Van Best and Van Dijk, 2020)
Figure 2.3 Illustration of major carbon sinks according to EPA 2020 classification
Tropical forests effectively sequester carbon because of their fast growth rates and high photosynthetic ability, which allow for substantial CO₂ absorption into living biomass such as leaves, stems, branches, and roots Their diverse biodiversity, with many different plant species, further enhances the ecosystems’ capacity to store and hold onto carbon Given the essential role tropical forests have in the global carbon cycle, measuring aboveground biomass (AGB) is crucial for estimating carbon capture potential and guiding efforts to achieve global carbon neutrality goals
2.2.2 Monitoring aboveground biomass in tropical forests
Aboveground biomass (AGB) describes the total mass of living organisms within a forest vegetation layer, including trees, shrubs, herbaceous plants, and other living organisms It offers a quantitative measure of vegetation mass per unit area and acts as a key metric in forestry for evaluating stand productivity and a forest’s carbon storage potential Monitoring AGB changes enables a better understanding of growth processes, ecological planning, and assessing impacts from natural disturbances and human activities This data supports sustainable forest management, biodiversity conservation, ecosystem restoration, and the efficient use of forest resources Since AGB is inherently linked to carbon sequestration and storage, it is central to climate change mitigation efforts The importance of accurate AGB estimation has been highlighted by international initiatives, especially within the REDD+ framework (Calvin et al., 2023; Moura et al., 2020; Srivastav, 2024) Among all forest carbon pools, AGB is the most dynamic component, responding quickly to changes in land management and forest use
The Sixth Assessment Report (AR6) of the Intergovernmental Panel on Climate Change (IPCC) emphasizes that information on biomass and carbon stock/sequestration at national and regional levels remains limited (Calvin et al.,
502 Bad GatewayUnable to reach the origin service The service may be down or it may not be responding to traffic from cloudflared
Methods for measuring biophysical features of trees
502 Bad GatewayUnable to reach the origin service The service may be down or it may not be responding to traffic from cloudflared
Figure 2.4 Illustration of tree components weighed separately
2.3.2 Measurement based on diameter at breast height and tree height
502 Bad GatewayUnable to reach the origin service The service may be down or it may not be responding to traffic from cloudflared
In tropical forest ecosystems characterized by high biodiversity, developing species-specific allometric models is often impractical due to limited data availability and high field survey costs Cumulative errors readily arise when input data are incomplete or inaccurate (T Ma et al., 2024) Therefore, many studies focus on group or ecosystem-level allometric models to balance accuracy and scalability Although specialized allometric equations for tropical forests have been created, they demand significant resources in terms of time and fieldwork Rugged terrain, dense vegetation, and limited infrastructure further hinder ground-based surveys Given these challenges, it is recommended to use allometric equations calibrated for regions with similar species composition to optimize resource use At the same time, remote sensing technologies such as optical, thermal, radar sensors, and LiDAR provide noninvasive methods for biomass data collection Satellite imagery and unmanned aerial vehicle (UAV) platforms allow for broader spatial and temporal coverage in biomass monitoring However, in-situ measurements are still essential for calibrating and validating multi-source remote sensing models to ensure accurate and reliable biomass estimates
Figure 2.5 Illustration of diameter at breast height and tree height measurement.
Remote sensing in tropical forest biomass estimation
Estimating aboveground biomass (AGB) is crucial for understanding the global carbon cycle and evaluating how forest ecosystems respond to climate change (T Ma et al., 2024) Rapid advances in remote sensing technologies over recent decades have improved the efficiency and spatial coverage of AGB assessments, from local to global levels Optical remote sensing, due to its freely available data and diverse spatial, spectral, and temporal resolutions, serves as a primary source for regional and national AGB analysis (Pilli et al., 2024; Ploton et al., 2017; Tian et al.,
2023) Additionally, commercial high-resolution satellite imagery has significantly improved the accuracy of biomass estimates in forested regions
Synthetic aperture radar (SAR) further extends AGB monitoring by providing medium to high spatial resolution observations regardless of weather conditions, enabling large-scale biomass estimation (Lourenỗo, 2021) A significant milestone in this field is ESA’s BIOMASS satellite, launched in April 2025 Operating in the P- band (wavelength ~70 cm), BIOMASS penetrates forest canopies to provide accurate measurements of tree height, forest structure, and biomass distribution, thereby supporting global estimates of forest carbon stocks (Le Toan et al., 2011; Liao et al., 2022; Quegan et al., 2019)
Figure 2.6 Illustration of remote sensing capabilities in forest biomass research 2.4.1 LiDAR in forest biomass estimation
LiDAR (Light Detection and Ranging) is an airborne laser scanning technology developed in the early 1970s It measures distances to targets using laser pulses to generate high-precision three-dimensional maps (accuracy < 1.0 m) Initially applied in topographic surveying, mapping, and geological research, LiDAR has, in recent decades, become a pivotal tool in forestry research and ecosystem monitoring Discrete-return LiDAR point clouds capture the three-dimensional structure of forest canopies, enabling precise determination of critical parameters such as tree height, canopy density, and biomass distribution at the plot level Its ability to penetrate foliage allows LiDAR to extract canopy metrics even in complex terrain where traditional ground-based methods are challenging Researchers integrate LiDAR data with field measurements to: (i) Estimate forest biomass at fine scales and monitor biomass dynamics over time; (ii) Classify species and derive individual tree biomass estimates at the stand level; (iii) Calibrate and enhance the accuracy of allometric models for aboveground biomass (AGB) estimation Owing to its high resolution and noninvasive data acquisition, LiDAR has markedly improved AGB estimation accuracy, thereby supporting forest conservation strategies, climate change assessments, and optimized forest resource management (Guo et al., 2023; Montesano et al., 2013)
Figure 2.7 Characteristics of LiDAR data for forest biomass estimation
LiDAR data can be collected from various platforms, including ground-based static systems, vehicle-mounted mobile units, unmanned aerial vehicles (UAVs), airborne sensors, and satellites Ground and mobile LiDAR systems produce high- density point clouds, allowing for accurate extraction of individual tree parameters such as diameter at breast height (DBH), tree height (H), and crown dimensions (Xiang et al., 2024) However, the high acquisition costs and large data volumes of these systems often restrict their use across extensive forest areas In contrast, airborne LiDAR provides three-dimensional forest structure information on large scales, but its spatial resolution is frequently too low for detailed biomass mapping Combining airborne LiDAR with other sources like optical remote sensing imagery and radar remains challenging due to differences in sensor characteristics and spatial resolutions Although LiDAR has been essential for estimating forest biomass and carbon stocks, expanding its use to landscape or regional levels is still limited by data acquisition costs and the complexity of data processing
2.4.2 Optical satellite imagery and synthetic aperture radar in aboveground biomass estimation
Multispectral optical satellite imagery is essential for estimating aboveground biomass (AGB) in tropical forests Optical sensors allow observations across various scales from global to local, with diverse spatial and temporal resolutions Coarse resolution systems like NOAA and MODIS provide frequent data, while Sentinel-2 MSI and Landsat-8/9 OLI offer moderate spatial detail Commercial satellites such as SPOT, Pléiades, WorldView, GeoEye, Ikonos, and QuickBird deliver high to very high resolution imagery, enabling detailed canopy cover analysis In biomass estimation, optical remote sensing indices, especially multispectral ones, are used to identify relationships between reflectance signals and biomass, primarily through canopy attributes like cover and foliar mass Optical reflectance generally shows a stronger correlation with aboveground biomass than with belowground biomass However, biomass estimation in tropical forest environments faces several challenges Local topographic effects, variable weather conditions, high biomass density, and the complex vertical structure of tropical canopies all affect signal quality and model accuracy (Kronseder et al., 2012; Phillips et al., 2019)
Advanced remote sensing techniques have been introduced and evaluated to improve the accuracy of tropical forest biomass estimation Vegetation-sensitive reflectance indices are recommended to reduce the effects of canopy shape, soil background, illumination angle, and atmospheric conditions on the measurement of plant physiological parameters (Abbas et al., 2020; Suwethaasri et al., 2025) The Landsat series, operational since the 1970s, has become a vital open-access dataset for passive remote sensing-based biomass assessment Alongside Sentinel-2, these platforms have enabled the development of various vegetation indices, including Normalized Difference Vegetation Index (NDVI), Wide Dynamic Range Vegetation Index (WDRVI), Enhanced Vegetation Index (EVI), and Leaf Area Index (LAI) All of these indices show strong correlations with biomass estimates (Goswami et al., 2015; Parker, 2020; Xue and Su, 2017), and regression studies have demonstrated their efficacy in predicting forest biomass (Ceccato et al., 2001; Xie et al., 2015) High-resolution imagery (< 5m) from SPOT-6/7, QuickBird, and WorldView enables detailed stand-level biomass modeling However, the high cost and limited availability of these commercial data pose challenges for their widespread use in AGB estimation
Passive remote sensing sensors are among the most common and practical tools for estimating aboveground biomass (AGB) Their benefits include a variety of spatial resolutions, high temporal reliability, global coverage, and relatively low acquisition costs (Abbas et al., 2020; Komiyama et al., 2008; Thi Nhung et al., 2024) However, because they depend on reflected light from the Earth’s surface, these sensors have limited penetration abilities and mainly capture horizontal vegetation structure This limitation hinders accurate representation of three-dimensional forest structure, including key parameters like tree height (H) and diameter at breast height (DBH), which are difficult to infer directly from passive optical data As a result, errors in AGB estimation can increase significantly (Ceccato et al., 2001; Y Ma et al., 2024) Moreover, signal saturation often occurs in densely vegetated areas, reducing the sensitivity of optical reflectance indices and impeding discrimination of high biomass values Therefore, despite their strengths in large-scale data acquisition, passive sensors pose significant challenges for AGB estimation in tropical forests
Synthetic aperture radar (SAR) is an active remote sensing system that sends microwave pulses and captures the backscattered echoes from target surfaces Unlike passive sensors, which depend solely on solar illumination, SAR can gather data and generate imagery continuously, both day and night, and is not affected by cloud cover or poor weather conditions Its capacity to penetrate vegetation canopies and clouds enables detailed analysis of subsurface forest structures Due to its reliable operation in various weather scenarios, SAR data offer a solid basis for estimating aboveground biomass (AGB) To estimate AGB, backscatter coefficients and polarization channels (HH, HV, VH, VV) are analyzed across multiple radar frequency bands to model how microwave signals interact with vegetation components The main bands include: (i) the short-wavelength X-band (∼3 cm), which mainly interacts with leaves and small canopy elements, providing detailed information about surface structures; (ii) the medium-wavelength C-band (∼5–6 cm), which penetrates the canopy and scatters off small branches and understory vegetation; (iii) the longer-wavelength L-band (∼23 cm), which detects responses from larger branches and trunks, aiding in modeling forest structure at intermediate depths; and (iv) the longest-wavelength P-band (∼70 cm), which passes through the entire canopy and produces backscatter mainly from tree trunks and the trunk-ground interface (Liao et al., 2022; Quegan et al., 2019)
Although synthetic aperture radar (SAR) can penetrate vegetation canopies to gather information on the vertical structure of forest stands and has been extensively used in biomass research, it still encounters several limitations in providing high- accuracy AGB estimates One main challenge is that SAR backscatter mainly reflects surface roughness, which makes it hard to distinguish between different vegetation types (Bui et al., 2024; Van Pham et al., 2019) Furthermore, SAR polarization signals are vulnerable to interference from environmental factors such as high wind speeds, humidity, and temperature fluctuations, which create unwanted variability in measurements and make biomass estimation more difficult Signal saturation in areas with dense biomass further reduces the accuracy of SAR-based analyses, especially in high-density forests, by limiting backscatter returns and decreasing sensitivity to biomass differences These limitations highlight the need for advanced analytical and modeling techniques to reduce noise effects and enhance the accuracy and reliability of SAR-derived forest biomass assessments
• The integration of synthetic aperture radar (SAR) and optical data
Most satellite-based sensors are affected by environmental factors (such as cloud cover, precipitation, and fog), sensor-specific characteristics (like spectral resolution, backscatter phenomena, and signal saturation), and terrain Among these, signal saturation is the leading cause of a significant reduction in image sensitivity when observing high-biomass tropical forests Numerous studies have shown that the saturation threshold depends on slope gradient, vegetation type, and the ecological attributes of the study area (Bui et al., 2024; Thi Nhung et al., 2024) Optical sensors, due to their limited penetration ability, often cannot gather information from understory layers in dense forests (Liao et al., 2022) Under closed canopy conditions, woody biomass keeps increasing while surface spectral reflectance stays nearly the same, leading to lower sensitivity and significant errors in aboveground biomass (AGB) estimation (Main et al., 2011; Wanyama et al., 2025) To overcome these limitations, integrating optical data with synthetic aperture radar (SAR) has become an effective solution SAR not only provides three-dimensional representations of forest structure from canopy foliage to trunks, but also operates reliably under all weather conditions, thereby improving the accuracy of AGB estimation and carbon stock assessment (Becker et al., 2023)
The integration of passive and active remote sensing data from multiple platforms generally yields more accurate aboveground biomass (AGB) estimates than relying on a single data source (Bui et al., 2024; Van Pham et al., 2019) This multisensor approach is critical in tropical forests, where complex environmental conditions and high biodiversity can reduce the effectiveness of traditional or single- sensor methods However, despite the potential of combining synthetic aperture radar (SAR) and optical imagery to enhance AGB retrievals, comprehensive studies on this combined approach remain limited for humid tropical ecosystems, especially within protected forest reserves In addition to selecting suitable remote sensing inputs, choosing the right estimation algorithms is crucial for developing reliable AGB estimation models Conventional statistical regressions are efficient and straightforward to run, but they cannot fully capture complex nonlinear relationships between biomass and sensor observations To address these limitations, machine learning methods such as Decision Trees (DT), Artificial Neural Networks (ANN), K-Nearest Neighbors (KNN), and Support Vector Machines (SVM) have been used to improve AGB estimation accuracy from remotely sensed data Many studies show that tree-based methods and Random Forests (RF) perform well in biomass modeling Additionally, the many tunable hyperparameters in these algorithms allow for customization to local study conditions, which further enhances model accuracy However, hyperparameter optimization and fine-tuning are often overlooked, even though parameter sensitivity can significantly influence the accuracy and reliability of SAR and optical-based AGB estimates
Figure 2.8 Illustration of the potential for forest ecosystem penetration from remote sensing data sources (passive and active) in forest biomass estimation
Based on a synthesis of previous studies and related literature, this research has compiled, selected, and developed a suite of remote‐sensing indices (optical and radar) alongside topographic variables for aboveground biomass estimation These indices are classified into groups based on their reflectance and backscatter mechanisms and their sensitivity to vegetation such as (i) Multispectral bands: Blue, Green, Red, Vegetation rededge, Nir infrared (NIR), and Shortwave Infrared (SWIR);
(ii) Ratio index: Ratio Vegetation Index (RVI), Redness Index (RI), Greenness Index
(GI), Normalized Difference Vegetation Index (NDVI), NDII Normalized Difference Infrared Index (NDII), Moisture Stress Index (MSI), Green Ratio Vegetation Index (GRVI), Chlorophyll Index Green (CIgreen), Chlorophyll Index Rededge (Cirededge), Aerosol Free Vegetation Index using SWIR-1 (AFRI-1), Aerosol Free Vegetation Index using SWIR-2 (AFRI-2), and Modified Simple Ratio (MSR); (iii)
Linear combinations index: Enhanced Vegetation Index (EVI), Soil Adjusted
Vegetation Index (SAVI), Normalized Difference Water Index (NDWI), Green Normalized Difference Water Index (GNDVI), Green Leaf Index (GLI), and Wide Dynamic Range Vegetation Index (WDRVI); (iv) Biophysical variables index: Leaf Area Index (LAI), Fractional vegetation cover (FVC), Leaf Chlorophyll Index (LCI), and Fraction Of Absorbed Photosynthetically Active Radiation (FAPAR); (v) Synthetic aperture radar index: Vertical Transmit-Vertical Receive Polarizations
(VV), Vertical Transmit-Horizontal Receive Polarissation (VH), Polarization mean (meanVHVV), Polarization multiply (multiVHVV), Square of a sum (sosVHVV), and Sum of cubes (socVHVV); (vi) Terrain index: Slope, Elevation, and Aspect Constructing this hierarchical framework of index groups is essential for standardizing inputs to biomass estimation models It provides a basis for analyzing, calibrating, and evaluating the suitability of each group under the specific topographic and ecological conditions of the study area In particular, the division of vegetation indices (VIs) into ratio-based and linear combination categories formalizes reflectance-based approaches to vegetation monitoring Ratio indices exploit simple relationships between two spectral bands to capture biomass variation and canopy greenness, offering straightforward implementation over large areas In contrast, linear combination indices derived using principal component analysis (PCA) integrate multiple spectral bands with weighted coefficients to minimize soil background effects, atmospheric disturbances, and sensor noise This classification not only clarifies the operational principles and application domains of each index type but also lays the foundation for evaluating index performance and developing advanced analytical tools in vegetation remote sensing
Machine learning models for estimating AGB in tropical forests
Globally, many methods and algorithms have been created to measure forest biomass (T Ma et al., 2024; Talebiesfandarani and Shamsoddini, 2022; Zhang et al.,
2020) Broadly, these methods fall into two categories: (i) Parametric and (ii) Nonparametric algorithms Parametric algorithms assume a predefined functional form, where a finite set of parameters describes the relationship between the dependent variable (biomass) and independent variables; univariate and multivariate linear regression are typical examples In contrast, nonparametric algorithms impose no prior assumptions about the mathematical function and instead build models directly from observational data, offering greater flexibility to capture complex, nonlinear relationships In the context of aboveground biomass estimation, machine learning (ML) techniques have become the dominant approach due to their ability to automatically extract features and provide higher predictive accuracy on new data Traditional regression models, most notably linear regression (LR), often show limited precision and poor generalization, which undermines their practical effectiveness To address these issues, recent research has increasingly used ML algorithms such as decision trees, random forests, and artificial neural networks to enhance both the accuracy and robustness of forest biomass estimates
Several widely used machine learning techniques have been employed for aboveground biomass estimation, including Random Forest Regression (RFR), Support Vector Regression (SVR), Artificial Neural Networks (ANN), Genetic Algorithms (GA), and Gradient Boosted Regression Trees (GBRT) These methods demonstrate strong ability in managing nonlinear, high‐complexity datasets and have enhanced AGB prediction accuracy compared to traditional regression methods (Bui et al., 2024; Thi Nhung et al., 2024) However, they still face significant challenges, mainly their heavy dependence on complex hyperparameter settings For example, choosing and tuning hyperparameters like the number of layers and hidden neurons in neural networks, decision-tree depth, learning rate, and loss function often requires extensive manual effort and expert knowledge Additionally, the interpretability of many machine learning models, especially deep convolutional architectures, remains limited, making it harder to explain and transparently justify their decision-making processes
Recent studies (Chahboun and Maaroufi, 2022; Do et al., 2023) have demonstrated the superior performance of the Cubist model, a rule-based inference algorithm for forest biomass estimation By integrating decision tree structures with multivariate linear regression, Cubist achieves both high predictive accuracy and model interpretability, while maintaining rapid computation This hybrid approach enables effective modeling of complex nonlinear relationships without excessive hyperparameter tuning, and it allows for quantitative assessment of each input variable’s contribution Compared to other machine learning techniques, Cubist has proven its superiority in accuracy, explanatory power, and computational efficiency (Campbell et al., 2024) These characteristics make Cubist an ideal choice for forest biomass research, especially when balancing predictive accuracy and transparency in decision-making Therefore, this thesis will use the Cubist model to develop and calibrate aboveground biomass estimates for tropical forests by combining multisource remote sensing data with topographic parameters.
STUDY AREA AND METHODOLOGY
Study area
3.1.1 Geographical location and tropical forest ecosystems
In Vietnam, forests are essential for maintaining ecosystem balance and significantly impact socio-economic development, especially in mountainous provinces where people rely mainly on integrated agricultural and forestry activities
In tropical forests, particularly in Vietnam's Northeast region, estimating total carbon stock using large-scale field measurements combined with remote sensing data remains challenging due to complex vegetation structures, varied terrain with steep slopes, and a lack of long-term datasets for research Bac Kan is located in a hilly and mountainous area, surrounded by rivers, and is one of the special provinces in the Northeastern part of Vietnam, with geographic coordinates ranging from 21°48' to 22°44' North latitude and 105°26' to 106°15' East longitude Administratively, it borders Lang Son province to the east, Tuyen Quang province to the west, Thai Nguyen province to the south, and Cao Bang province to the north Bac Kan comprises 08 administrative units, including Bac Kan city and seven districts: Bach Thong, Ba Be, Cho Don, Cho Moi, Na Ri, Ngan Son, and Pac Nam (Figure 3.1) Bac Kan province, which has the highest forest cover rate in Vietnam, plays a vital role in absorbing and storing carbon, thereby making significant contributions to efforts to reduce greenhouse gas emissions Accurate assessment and quantification of accumulated carbon stock are crucial, especially given the rapidly growing forest carbon credit market that offers substantial potential to support sustainable forest ecosystem conservation and development However, the lack of comprehensive biomass data for the province’s forested areas poses major challenges in calculating carbon credits for emission reduction policies, which diminishes the overall effectiveness of their implementation This data gap also hinders the fulfillment of multilateral agreements and climate change mitigation commitments at both national and local levels in Vietnam In high-value forest ecosystems, preventing degradation and conserving areas with significant carbon stocks are essential These actions not only help reduce the effects of climate change but also protect the environment, maintain ecosystem functions, and preserve the genetic diversity of plants and animals
Bac Kan province, with its rich forest ecosystems, makes an essential contribution to biodiversity conservation and climate change mitigation Thirteen main forest ecosystem types of Bac Kan province have been synthesized, including: Evergreen broadleaved forest on soil mountain (EBFS), Evergreen broadleaved forest on rocky mountain (EBFR), Mixed broadleaved and coniferous forest on soil mountain (MBCRS), Mixed wood-bamboo forest on soil mountain (MWBFS); Indosasa bamboo forest on soil mountain (IBFS), Neohouzeana bamboo forest on soil mountain (NBFS), Palm and coconut tree forest on soil mountain (PCTFS), Palm and coconut tree plantation on soil mountain (PCTPS), Plantation on soil mountain (PFS), Plantation on rocky mountain (PFR), Bamboo plantation on soil mountain (BPS), Bamboo plantation on rocky mountain (BPR), and Other plantation on soil mountain (OPFS) Forest ecosystems in Bac Kan province were identified and classified using the 2020 forest dynamics inventory data compiled by the Forest Inventory and Planning Institute (FIPI) under the Ministry of Agriculture and Environment, and the study area comprises three special use forest units: (i) Ba Be National Park covers over 10,000 hectares; (ii) Kim Hy Nature Reserve has a core zone exceeding 15,700 hectares and a buffer zone of nearly 23,000 hectares; and (iii) Nam Xuan Lac Species and Habitat Conservation Area features a 4,100-hectare core zone and a 16,300- hectare buffer zone The vegetation assemblages within these forests are exceptionally diverse and of high ecological value, serving as key conservation centers for the genetic resources of flora and fauna in Vietnam’s northeastern region
3.1.2 Analysis of natural and socio-economic characteristics of Bac Kan province
• Terrain characteristics of the study area
Bac Kan is a mountainous province located in the center of the Northeast region of Vietnam, with an area of about 4,860 km² About 80% of the land is mountainous, with an average elevation of around 600 meters above sea level The terrain features high mountain ranges and steep slopes, forming a varied landscape of valleys, low hills, medium mountains, and high peaks Geomorphologically, the province is classified into four main types: (i) high mountains in the western and northern parts; (ii) low mountains formed on terrigenous sedimentary rocks; (iii) limestone mountains associated with the Ngan Son arc; and (iv) valleys shaped by tectonic and erosional processes In terms of elevation, the province is divided into four levels: (1) areas below 200 meters, mainly in the center and along the Cau and Nang river basins; (2) from 200 to 600 meters, covering most of the area with slopes of 25° to 30°; (3) from 600 to 1,000 meters, with slopes of 25° to 40°; and (4) above 1,200 meters, including peaks with slopes exceeding 40° (see Figure 3.2) The region's diverse topography, significant dissection, and an average slope of 26° support rich ecosystems and abundant natural resources However, these features also present challenges for forest management and conservation efforts
• Climatic characteristics and hydrological system
Bac Kan province is located in the tropical monsoon belt of Southeast Asia, with a climate that shifts between cold, dry winters and hot, rainy summers The rainy season lasts from April to October, while the dry season extends from November to March of the following year Monsoons are the main factor affecting seasonal rainfall patterns and the distribution of precipitation over time and space The province receives an average annual rainfall of about 1,084 mm, though this varies across regions and seasons The average humidity is around 84%, and temperatures typically range from 23 to 24°C Bac Kan features a relatively dense and evenly distributed hydrographic network, including numerous rivers and streams Notably, five major rivers in the northeastern region are the Nang, Cau, Pho Day, Bac Giang, and Na Ri rivers During the rainy season, water from mountain slopes quickly flows into narrow valleys, causing river and stream levels to rise rapidly and leading to flooding in low-lying areas Conversely, during the dry season, river discharge drops sharply, and in the driest months, many rivers and streams flow weakly or are nearly dry The seasonal variation in flow is obvious, with most rivers and streams heavily influenced by rainfall and flood events These climate and hydrological features are essential for managing water resources and maintaining forest ecosystems in Bac Kan province
Figure 3.2 Spatial distribution of elevation (left) and slope (right) in Bac Kan province
• Socio-economic characteristics and position of Bac Kan province
Bac Kan is a mountainous province located in the central part of northeast Vietnam, playing an important role in socio-economic development The province covers approximately 4,860 km², with a population of about 330,000 and a density of roughly 68 people per square kilometer Of this population, around 80,000 live in urban areas, while approximately 80,000 reside in rural areas The demographic composition of Bac Kan province is predominantly made up of ethnic minorities, who constitute 87% of the total Among these, the Tay ethnic group has the most significant proportion (52%), followed by the Dao (18%), Nung (9%), H'mong (7%), and San Chay (1%) This ethnic diversity has fostered a rich cultural landscape, with distinct customs and traditions specific to each group, while also exerting a profound influence on the province’s socio-economic development Developing programs and policies tailored to different ethnic communities, preserving traditional cultural values, and integrating sustainable economic development strategies have become both challenges and opportunities for Bac Kan
The province has significant potential for developing forest-based agricultural and forestry industries, mainly due to its rich forest resources According to statistics, Bac Kan province's forested land area, including both unclosed canopy plantations, covers up to 374,027.12 hectares, with natural forests making up about 271,805 hectares and planted forests covering 102,222 hectares In 2023, the forest coverage rate reached 73.38%, showing Bac Kan’s substantial potential for forest resource utilization and growth By leveraging these advantages, the province has prioritized sustainable forest-based economic development as a crucial sector that contributes significantly to socio-economic progress The forestry sector not only produces high monetary value but also provides stable jobs and supports sustainable poverty reduction for rural communities, predominantly ethnic minority groups that rely on forests Programs and projects focused on sustainable forest development, better forest resource management, and promoting forest-based products such as timber, medicinal plants, and non-timber forest items will play a vital role in improving local livelihoods and enhancing the province’s prosperity economy
Therefore, with its natural conditions, socio-economic characteristics, and the province's strengths in forests and forest land, Bac Kan functions as an enormous carbon reservoir and is poised to benefit significantly from participating in the global carbon credit market
3.2 Multi-temporal satellite imagery and GIS database
Sentinel-2 MSI was launched into orbit on June 23, 2015, at an approximate altitude of 786 km, with the mission to capture high-resolution multispectral images, including a total of 13 spectral bands and a swath width of 290 km Meanwhile, synthetic aperture radar (SAR) is a radar technique used to produce high-spatial- resolution imagery by transmitting electromagnetic waves toward the Earth’s surface and recording the backscattered signals from surface objects Sentinel-1A, equipped with C-band SAR, was launched into orbit on April 3, 2014, at about 693 km altitude
It can penetrate vegetation layers but cannot fully pass through dense materials such as thick tree trunks As a result, C-band radar provides valuable information about canopy structure and the upper layers of vegetation, aiding in the analysis and estimation of forest aboveground biomass (AGB) under all weather conditions and offering significant potential for determining stored carbon
However, the integration of Sentinel-1 SAR and Sentinel-2 MSI data has not been extensively explored in AGB estimation models within tropical rainforest ecosystems, especially in specialized forests Consequently, this study uses Sentinel-
2 MSI and Sentinel-1A SAR satellite data from 2016, 2020, and 2025 to calculate multi-temporal AGB in forest ecosystems of Bac Kan province, Vietnam Specifically, for Sentinel-2 MSI images with under 2% cloud cover, data were freely accessed from (https://dataspace.copernicus.eu), and Sentinel-1A SAR images were obtained from (https://search.asf.alaska.edu) The imagery scenes within the study area were selected to be as close in acquisition time as possible, facilitating remote sensing data preprocessing The selection criteria included: (i) Ensuring optimal image quality compared to other images from the same period; (ii) Using imagery from a specific date to maintain consistency and reduce the influence of temporary factors such as lighting conditions, cloud cover, and noise, which, although they do not significantly affect forest area, may impact image quality; and (iii) Avoiding the need for extensive time and resources that might be required when processing and analyzing multiple images captured at different times Therefore, this study selected a set of multi-source remote sensing datasets with acquisition times close together to ensure the best possible image quality These datasets were integrated with ancillary data, such as survey maps and forest inventory maps, to compile and classify characteristic forest ecosystems within the study area The 1:50,000-scale GIS database was obtained from the project code ĐTĐL.CN-42/23 and the Department of Agriculture and Environment of Bac Kan province, while the forest inventory map was collected from the Forest Inventory and Planning Institute (FIPI)
Table 3.1 Data used and characteristics of Sentinel-1 SAR and Sentinel-2 MSI
Sentinel-2 MSI (Level-1 data) Multispectral
This study uses a field dataset of 251 sample plots, ensuring representativeness for AGB in both natural and planted forests The sample plots were collected, updated, and calculated from 2020 to 2025 Each plot was selected based on these criteria: i) representing the forest types studied and their topographical conditions; ii) including various tree size classes; iii) located in minimally disturbed forests where large trees are present For the natural forest survey plot, which covers 1,000 m², a circular design was used with a radius of 17.84 m, centered on the plot's grid point (located within an 8 km × 8 km area), with the distance between the centers of neighboring plots being 150 m (Figure 3.3)
Figure 3.3 Illustration of a cluster of sample plots in Bac Kan Province
Data collection was carried out in natural forests, including evergreen broadleaf forests on low mountains, rocky mountains, mixed bamboo-wood forests on low mountains, bamboo forests on low mountains, and coconut palm forests on low mountains For the planted forest sample plots, each covered an area of 500 m², was circular with a radius of 12.62 meters, and had its center aligned with the measurement point These plots aimed to gather data on plantation trees, both on soil mountains and rocky mountains For trees with a diameter at breast height (DBH) of
1.3 meters or more and more than 6 cm, detailed counts and measurements were performed using distance-measuring devices, height gauges, and self-leveling tools For trees with a DBH less than 6 cm, only the number of trees within the sub-sample plots was recorded Bamboo plots, each with an area of 100 m², were circular with a radius of 5.64 meters, and their centers matched the measurement points These plots focused on surveying bamboo in natural forests, non-forested areas, and upland bamboo plantations The key parameters from these plots were used to estimate actual AGB by applying biomass equations developed explicitly for the current forest ecosystem in Bac Kan (Huy et al., 2016)
Figure 3.4 Illustration of inventory measurement activities in a tropical forest ecosystem
In evergreen forest ecosystems, including Evergreen broadleaved forest on soil mountain, Evergreen broadleaved forest on rocky mountain, and Mixed broadleaved and coniferous forest on soil mountain, the allometric equation was applied using the formula (1) This formula, which relates tree measurements to biomass, is a standard method for estimating AGB For other forest types, including mixed wood-bamboo forest on soil mountain, Indosasa bamboo forest on soil mountain, Neohouzeana bamboo forest on soil mountain, Palm and coconut tree forest on soil mountain, Palm and coconut tree plantation on soil mountain, Bamboo plantation on soil mountain, and Bamboo plantation on rocky mountain, a different allometric equation was applied using the formula (2) Similarly, for plantation forest ecosystems, including Plantation on soil mountain, Plantation on rocky mountain, and Other plantation on soil mountain, a specific allometric equation was applied using the formula (3)
Aboveground biomass (AGB) values ranged from 15.74 Mg ha⁻ạ to 361.67
The methodology framework used in the study
The study uses a comprehensive approach based on a modern methodological framework that combines multi-source remote sensing data (including multispectral and radar satellite images), topographic information, and field measurement data to estimate aboveground biomass (AGB) and cumulative carbon stock (CS) in Bac Kan province The research process is divided into four main steps, shown in Figure 3.5
The first step involved integrating and processing multispectral, radar, and topographic satellite data to derive key indices for the AGB estimation model In the second step, field forest inventory data were analyzed, and allometric equations were used to calculate actual biomass The third step focused on developing a Cubist machine learning model using the processed indices to predict AGB across the study area In the final step, spatial analysis of the predicted AGB was conducted to assess the biomass distribution and carbon accumulation potential of high conservation value tropical rainforest areas in Bac Kan Details of each step are explained in the following sections.
Figure 3.5 Flowchart of the research procedure for assessing tropical forest biomass stocks and cumulative carbon in Bac Kan province.
3.3.1 Processing data and identification of selection indices
Remote sensing technology has dramatically improved the effectiveness of vegetation monitoring Optical vegetation indices (OVIs), which are calculated using simple mathematical formulas or reflectance transformations in two or more spectral channels across different wavelengths, are used to represent the condition of vegetation while minimizing interference from other source factors OVIs can reduce or eliminate the effect of surface features and atmospheric conditions, enabling a focus on vegetation-specific aspects of information Therefore, OVIs have been widely used in numerous studies ranging from local to global scales and across most fields of Earth Science, especially in forest ecology research The optical vegetation indices are based on the physical principles of spectral reflectance captured by remote sensors, reflecting the complex interactions among light, the atmosphere, and vegetation The spectral reflectance recorded by satellite sensors from solar radiation is a mixed signal, including atmospheric noise, reflectance from leaves, canopies, plant structures, shadows, soil, and other ground factors, as well as from the satellite
Therefore, spectral radiation calibration is essential for removing noise and accurately capturing the actual spectral reflectance of the surface object (Pham et al.,
2021) This process improves the accuracy of both qualitative and quantitative analysis of forest biomass In this study, atmospheric and topographic correction (ATCOR), integrated within the Catalyst Professional software package (https://catalyst.earth/), was used to correct radiometric and atmospheric effects on surface reflectance The ATCOR workflow included: (i) Top-of-atmosphere (TOA) reflectance correction to remove the influence of solar radiation and atmospheric conditions; (ii) Haze removal to eliminate thin clouds and improve the clarity and accuracy of reflectance data; and (iii) Converting radiance values to surface reflectance to accurately depict the optical properties of vegetation (Figure 3.6) Applying the ATCOR method not only enhances the reliability of optical remote sensing data but also enables more accurate modeling and estimation of aboveground biomass (AGB)
Figure 3.6 Satellite imagery processed through atmospheric correction and converted from radiance to surface reflectance for the period 2016–2025
Synthetic aperture radar (SAR) is an active remote sensing system that works by transmitting signals and receiving the backscattered responses from surface objects Unlike passive sensors, SAR can operate effectively both day and night, independent of solar radiation, and can penetrate clouds and vegetation, enabling the collection of information on forest structure characteristics In particular, SAR maintains stable performance under all weather conditions, supporting aboveground biomass estimation through radar backscatter coefficients However, C-band SAR data are generally reliable, but they can still be affected by soil moisture and vegetation water content, leading to variability in backscatter responses To reduce these impacts, Sentinel-1A SAR satellite images with dual polarizations, VV (Vertical-Vertical) and VH (Vertical-Horizontal), are processed using the Sentinel-1 Toolbox (S1TBX) within the open-source software package SNAP
(https://step.esa.int/) Perform radiometric calibration to convert raw image digital number (DN) values into radar backscatter coefficients (sigma θ), ensuring data consistency; reduce speckle noise using the enhanced Frost adaptive filter with a 7×7 pixel moving window to smooth the backscatter coefficients (Van Pham et al., 2019), Terrain correction using a digital elevation model (DEM) with a spatial resolution of
10 m to improve the geometric accuracy of SAR backscatter responses (Figure 3.7)
Figure 3.7 SAR data after preprocessing, converting DN values to backscatter (dB) for VH and VV polarizations for the period 2016-2025
Topographic indices are essential for estimating tropical forest biomass because they directly influence environmental factors that affect vegetation growth The slope index impacts soil drainage and water retention Areas with steep slopes tend to drain water quickly, leading to lower soil moisture and hindering tree growth, which can reduce biomass Conversely, areas with gentle slopes help retain water longer, creating favorable conditions for tree growth and biomass accumulation This is also a critical coefficient when using field-based methods to calculate aboveground biomass that needs to be collected The elevation index significantly impacts climate and temperature conditions Higher elevations usually have harsher climates and cooler temperatures, which restrict vegetation growth and result in less biomass In contrast, lower elevations typically have warmer climates and higher humidity, creating ideal conditions for trees to grow and accumulate biomass The aspect index affects how much solar radiation an area receives Areas facing the sun (south-facing in the northern hemisphere) experience higher temperatures and lower moisture levels, while shaded areas (north-facing in the northern hemisphere) tend to be cooler and wetter, supporting the growth of particular tree species and increasing biomass
When estimating aboveground biomass (AGB) using machine learning models, including too many remote sensing–derived indices can unnecessarily increase model complexity It may also cause multicollinearity among input variables, decrease model stability, and extend processing time These effects can ultimately reduce the accuracy of biomass predictions To improve the research workflow, an important step is to remove indices with minimal contributions Correlation analysis is used to examine the relationships among indices, while the influence of each index on the AGB estimation model is evaluated to identify the most relevant set of predictors Selecting an optimal subset of indices not only improves model performance but also ensures practicality in real-world applications
In this study, experimental analysis was conducted to choose indices derived from optical, radar, and topographic remote sensing data A total of 251 sample plots served as the basis for assessing correlations among these indices The results showed that all indices were statistically significant (p< 0.01), with correlation coefficients calculated within and between index groups
(i) Within the group of single-band spectral indices, analysis results showed linear correlations among the index (Figure 3.8) However, the standard distribution plots of the NIR, GREEN, SWIR-1, REDEDGE-1, and REDEDGE-4 bands demonstrated better performance compared to the other single-band spectral channels
(BLUE, RED, REDEDGE-2, REDEDGE-3, SWIR-2) Therefore, five indices from these single-band spectral channels, including NIR, GREEN, SWIR-1, REDEDGE-
1, and REDEDGE-4, were selected for use in the AGB estimation model
Figure 3.8 Correlation between single-band spectral indices from Sentinel-2 MSI
(ii) Among the ratio indices: RVI, RI, GI, NDVI, NDII, MSI, GRVI, CIGREEN, CIREDEDGE, AFRI-1, AFRI-2, and MSR, Figure 3.9 shows the positive and negative correlations between these indices The analysis results indicated that MSR, AFRI-1, AFRI-2, CIGREEN, GRVI, NDII, NDVI, and CIREDEDGE exhibited more favorable distribution patterns compared to RVI, RI, GI, and MSI However, there was a perfect positive correlation between the index pairs (NDVI and MSR), (AFRI-1 and AFRI-2), and (CIGREEN and GRVI) Among these, the standard distribution plots of MSR, GRVI, and AFRI-2 show better performance compared to NDVI, CIGREEN, and AFRI-1 Therefore, MSR, AFRI-2, GRVI, NDII, and CIREDEDGE were selected to participate in the model for estimating aboveground biomass
Figure 3.9 Correlation between ratio-based indices from Sentinel-2 MSI
(iii) Within the group of linear combination indices, EVI, SAVI, NDWI, GNDVI, GLI, WDRVI, and the group of physiological variables, LAI, FVC, LCI, FAPAR Figure 3.10 shows the correlation between the indices in each group The analysis results indicate that EVI, SAVI, NDWI, GNDVI, GLI, and WDRVI all follow a normal distribution, accurately reflecting the actual data However, in the group of linear combination indices, there are two pairs of indices (EVI and SAVI),
(NDWI and GNDVI) that show both positive and negative correlations Meanwhile, the standard distribution plots of EVI and NDWI were more optimal than those of SAVI and GNDVI, indicating a better ability to reflect actual data Therefore, EVI, NDWI, GLI, and WDRVI were selected to participate in the model for estimating aboveground biomass Similarly, the group of physiological indices, including LAI, FVC, LCI, and FAPAR, also follows a normal distribution, ensuring reliability in data analysis Thus, all indices in this group were included in the AGB estimation model
Figure 3.10 Correlation among indices in the linear combination and vegetation physiology groups from Sentinel-2 MSI
(iv) In the two groups of synthetic aperture radar indices, including VV, VH, meanVHVV, multiVHVV, sosVHVV, and socVHVV, as well as topographic indices, including SLOPE, ELEVATION, and ASPECT, the standard distribution plots of both groups are shown in Figure 3.11 The analysis indicates that VV, VH, meanVHVV, multiVHVV, and sosVHVV follow a normal distribution However, multiVHVV and sosVHVV are perfectly positively correlated, with sosVHVV exhibiting a more ideal normal distribution compared to multiVHVV For the topographic indices, both SLOPE and ELEVATION also follow a normal distribution, ensuring reliability when used in the model Based on these results, the indices selected for the AGB estimation model include: VV, VH, meanVHVV, sosVHVV, SLOPE, and ELEVATION
Figure 3.11 Correlation between synthetic aperture radar indices and topography
Based on the correlation analysis among groups of indicators, including single-band spectral indices, ratio-based indices, linear combination indices, physiological vegetation variables, SAR indices, and topographic indices, the study has selected 24 appropriate indices (Table 3.2) These indices include NIR, GREEN, SWIR-1, REDEDGE-1, REDEDGE-4, MSR, AFRI-2, GRVI, NDII, CIREDEDGE, EVI, NDWI, GLI, WDRVI, LAI, FVC, LCI, FAPAR, VV, VH, meanVHVV, sosVHVV, SLOPE, and ELEVATION The chosen indices function as independent variables, incorporated into the AGB estimation model for the study area
Table 3.2 Selected indices from multi-source remote sensing and topographic data
Independent variables Name/Wavelength/Formula
Band 8 NIR-INFRARED (784 nm – 900 nm)
Band 11 SHORTWAVE INFRARED (1.565 nm – 1.655 nm)
SP E C T R AL I NDI C E S FR OM SENT INE L -2A
NDII = (NIR − SWIR1) (NIR + SWIR1)
(Green Ratio Vegetation Index) GRVI = NIR
(Aerosol free vegetation index) AFRI = (NIR − 0.5 × SWIR2
(Modified Simple Ratio) MSR = ( NIR
(Enhanced Vegetation Index) EVI = 2.5 × (NIR − RED)
(Green leaf index) GLI = (2 × GREEN − RED − BLUE)
Independent variables Name/Wavelength/Formula
(Leaf area index) LAI = (3.618 × EVI − 0.118)
(Fractional vegetation cover) FVC = (NDVI − NDVImin)
(Leaf Chlorophyll Index) LCI = (NIR − REDEDGE1)
(fraction of absorbed photosynthetically active radiation)
1 − NDVI) min ] + 0.001) + ( 0.949 × (NDVI − NDVI min )
Synthetic aperture radar index (SAR)
(Polarization mean) meanVHVV = (VH + VV)
(Square of a Sum) sosVHVV = VH 2 + 2 × (𝑉𝐻 × 𝑉𝑉) + VV 2
3.3.2 Cubist machine learning model for predicting aboveground biomass