Forest Aboveground Biomass Estimation and Inventory: Evaluating Remote Sensing-Based Approaches

Abstract

Remote sensing datasets offer robust approaches for gaining reliable insights into forest ecosystems. Despite numerous studies reviewing forest aboveground biomass estimation using remote sensing approaches, a comprehensive synthesis of synergetic integration methods to map and estimate forest AGB is still needed. This article reviews the integrated remote sensing approaches and discusses significant advances in estimating the AGB from space- and airborne sensors. This review covers the research articles published during 2015–2023 to ascertain recent developments. A total of 98 peer-reviewed journal articles were selected under the Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) guidelines. Among the scrutinized studies, 54 were relevant to spaceborne, 22 to airborne, and 22 to space- and airborne datasets. Among the empirical models used, random forest regression model accounted for the most articles (32). The highest number of articles utilizing integrated dataset approaches originated from China (24), followed by the USA (15). Among the space- and airborne datasets, Sentinel-1 and 2, Landsat, GEDI, and Airborne LiDAR datasets were widely employed with parameters that encompassed tree height, canopy cover, and vegetation indices. The results of co-citation analysis were also determined to be relevant to the objectives of this review. This review focuses on dataset integration with empirical models and provides insights into the accuracy and reliability of studies on AGB estimation modeling.

1. Introduction

Forest aboveground biomass (AGB) estimation and mapping is pivotal in understanding and managing carbon stocks and contributing to global efforts in climate change mitigation. Despite its significance, achieving accurate, large-scale assessments has remained challenging within the scientific community [1,2,3]. The tropics have four times larger gross emissions and removals than temperate and boreal ecosystems [4]. Deforestation is the main contributor to global forest carbon loss, which indirectly reduces the carbon absorption capacity of land and forest ecosystems [5]. The largest carbon pool is aboveground biomass (AGB), which varies widely depending on disturbance. Characterizing the Earth’s climate system and fighting climate change requires accurate AGB estimation, a vital climate variable [6]. This creates a need for remote sensing datasets, integrated with cost- and time-efficient approaches, to map and estimate forest aboveground biomass.

Advances in remote sensing using freely available datasets offer new opportunities for monitoring large-scale forest ecosystems (medium spatial resolution images) [7,8,9]. They have made it possible to monitor and map forest cover and changes over large coverage at regular intervals with reasonable accuracy [10]. Medium spatial resolution images are familiar sources for regional biomass estimation. Since coarser resolution data offer superior exchanges between temporal and geographical resolution and coverage, they are frequently utilized in analyzing forest AGB at the regional and global levels [11]. Active and passive remote sensing techniques make the estimation more accurate and time-efficient at local and regional levels [2]. To some extent, spaceborne Light Detection and Ranging, together with radar sensors and unmanned aerial vehicles, permits the determination of emission parameters previously obtained from field inventory data [3]. The multiple satellite observations of SAR backscatter provide new insights into terrestrial woody AGB’s spatial distribution and magnitude [12]. The integration approaches of SAR datasets with field inventory data are utilized to estimate the forest growing stock and assess its relative changes [13]. More recently, large-scale methods to map AGB have generally combined datasets from different sensors to overcome their limitations [10]. Integration of optical (passive) data with LiDAR (active) has produced more accurate estimates of the AGB [14]. Spaceborne remote sensing is currently the most helpful method for monitoring the AGB globally [15]. Active remote sensing techniques, such as LiDAR and SAR systems, are more sensitive to the vegetation structure than optical sensors [16]. The most accurate way to measure a forest’s vertical structure is with LiDAR, which can also increase the accuracy of estimating the forest’s aboveground biomass when combined with some optical images [17]. It tracks disturbance and regeneration in boreal, temperate, and tropical forests and provides insights into forest structure characteristics [18]. The other dataset is a spaceborne LiDAR imager, the Global Ecosystem Dynamics Investigation (GEDI), that estimates aboveground biomass and provides accurate information about forest structure [19]. In this case, obtaining comprehensive mapping of forest stand volume using GEDI data without optical sensor images will still be difficult because of spatial and temporal limitations [20]. GEDI products’ height profile and AGB density value were fused with optical datasets (Landsat-8) to classify the forest type in Mediterranean surroundings [21]. Empirical models utilized globally distributed coincident fields and airborne LiDAR datasets to create the GEDI-L4A (25 m footprint) product [22]. These sensors have been used to overcome and reduce individual limits in biomass estimation due to the low sensitivity of the biomass of forest components. These limitations have resulted in everyday issues such as saturation and partial observation [23]. Many modeling techniques were used to minimize the AGB estimate uncertainties at multiple spatial scales [24]. Rapid-Eye satellite images were employed with a random forest model and k-nearest neighbor method for tree species prediction and biomass mapping [25]. Synthetic aperture radar and multi-spectral imagery data using non-linear regression analysis were used to map AGB [26]. A random forest model and spaceborne L-band microwave vegetation opacity measurement were used to map changes in aboveground biomass across climatically driven events [27]. The integration of optical datasets with field plot data was used to estimate AGB using K-nearest neighbor imputation models [28]. The stock volume was assessed using the values of Sentinel-2 imagery and land-cover classification based on the same image, which was learned using the MODIS data following the empirical approach [29].

Previously, many studies have attempted to review the assessment of forest aboveground biomass by remote sensing techniques. Ahmad et al. [14] showed the importance of high-resolution optical satellites by estimating AGB using linear and non-linear regression with satellite-derived independent variables. Gyamfi-Ampadu E et al. [8] describes how remote sensing applications monitor tropical and sub-tropical natural forests, emphasizing change detection, AGB, tree species diversity, and their identification at the regional scale. The approaches of satellite remote sensing to assess AGB from national to global tropical forests are explained by Abbas S. et al. [1]. Turton A. et al. [30] documented statistical methods to improve forest aboveground biomass estimates and their changes. Mutanga O. et al. [31] attributed the spectrum saturation situations observed from remote sensing datasets to the high density of vegetation globally. However, synergetic integration methodologies for mapping and estimating forest AGB have not yet been synthesized in these evaluations.

This paper reviews the outcomes of synergistic approaches using diverse platform datasets for forest AGB mapping and estimation through remote sensing-based technologies. It gives a comprehensive insight into integrating such datasets, pointing out possibilities for further research and identifying gaps in order to improve current AGB mapping and estimate procedures, primarily focusing on information extraction and remote sensing techniques.

The approaches and workflows for estimating forest aboveground biomass using multiple remote sensing platforms can be segmented into three primary categories. First, spaceborne platform datasets have the potential to identify forest land on a larger footprint to estimate biomass using its spectral, textural, and height information. Second, airborne platform datasets can provide high spatial resolution data to estimate biomass on a local scale. Lastly, space- and airborne platform datasets have the potential for a process-based understanding to develop a synergic workflow for local- to large-scale AGB estimation and mapping.

Consequently, this research aims to (a) systematically review and analyze the effectiveness of multi-platform integration approaches in forest AGB estimation and (b) identify the most commonly used remote sensing datasets and regression models in these integrated approaches. Through this analysis, this study aims to provide insightful information about the current state of AGB mapping, assess the lessons learned from past research, and propose a future agenda to guide ongoing and future studies in this field.

2. Materials and Methods

The analysis of a systematic literature review (SLR) is a useful way of assessing ideas or indications within a specific field and determining whether or not individual hypotheses are accurate and powerful. The PRISMA framework was consulted in preparation for determining the specifics of this review process. Section 2.4 indicates that these papers were carefully appraised and filtered using previously established quality criteria. The fundamental PRISMA guidelines were followed in executing and publishing the review results. The complete workflow of this review is explained as shown in Figure 1.

Figure 1. Workflow of literature review. Grey color represents the processing steps, yellow color represents the article selection criteria, and orange represents discussion and analysis.

2.1. Search Approach and Terms

To attain the objectives of this review, a Boolean operator known as “AND” was combined with three keywords: “Mapping, Forest Aboveground Biomass Estimation, and Remote Sensing”. This set of keywords was applied as a query string in the selected databases. These three keywords helped make the published articles as accessible to review as possible. It was desired that all the published research articles, mostly in English, be included in this SLR so that no research papers in the given time frame were overlooked.

2.2. Timeline and Sources of Data

We searched through all mentioned digital data sources from 2015 to 2023, considering the current technology developments in forest aboveground biomass calculation and remote sensing. Since synergistic remote sensing platform data like Landsat-8, Sentinel-2, and GEDI datasets are a current focus of interest for researchers and would help highlight recent developments in this sector, the short period from 2015 to 2023 was selected We created a search strategy to customize this systematic search method to find pertinent literature. This search method was customized for five databases (Scopus, Science Direct, Nature, IEEE Xplore, and MDPI) to thoroughly locate pertinent papers for this review.

2.3. Selection Process

The selection criteria for research articles were based on the PRISMA process flow to categorize the relevant articles. This process is presented by Kitchenham [32] and includes four steps. Main research articles from sources were collected in the first step, along with identifying and removing duplicated articles from the databases. The second step was to screen the collected articles based on titles and abstracts to check their relevance to the review article objective. The third stage entailed comparing the articles’ quality standards to determine whether they were to be included or excluded, determining their eligibility. The final stage involved adding up all the research articles for quantity synthesis.

Using the abovementioned process, 1076 studies were found from the abovementioned databases. This was because most of the research articles used the keywords of “mapping”, “forest aboveground biomass estimation”, and “remote sensing”. These keywords are used in many articles available in the ScienceDirect database (812) that do not have any relevance to forest biomass estimation, e.g., [33,34,35], which were not included. Similarly, soil-related articles, e.g., [36,37], and even those with carbon-pool-related applications, were excluded. Among all the databases, 65 duplicated articles were also removed at this stage. After that, the search was narrowed by reviewing and removing publications without a quality assessment based on the exclusion criteria for mapping and estimating aboveground biomass in forests. At this stage, 245 research articles had filtered through. For checking the quality of articles as per the inclusion criteria, among all the databases, 98 publications had information which was included in the review of this article. These articles were further classified into three categories: spaceborne, airborne, and space- and airborne platforms. This workflow is shown in a flowchart that summarizes this article’s screening method in Figure 2.

Figure 2. Flowchart of article screening as per selection criteria.

We extracted the year of publication of each article, the details of the dataset, the type of forest, field data, its source, the region of study, and the accuracies for the models of forest estimation. Table 1 depict the number of scrutinized articles by year and database.

Table 1. Database-wise distribution of the articles and their allocation in the three categories.

Database Source	Spaceborne Platform	Airborne Platform	Spaceborne and Airborne Platform	Total
ScienceDirect	30	13	13	56
MDPI	10	4	3	17
Nature	1	1	0	2
Scopus	10	2	4	16
IEEE Xplore	3	2	2	7
Total	54	22	22	98

2.4. Inclusion and Exclusion Criteria

This section discusses and covers the eligibility criteria that have been formulated and implemented in place regarding the research articles scrutiny. As per the developed criteria, research articles from the mentioned source journals were included from 2015 to 2023. The articles were selected based on titles and abstracts that mainly mentioned mapping and estimating forest aboveground biomass. The detailed process of the article selection was discussed in Section 2.1, Section 2.2 and Section 2.3. This systematic literature review (SLR) excluded non-English articles and those lacking remote sensing for mapping and forest aboveground biomass estimation. Similarly, other studies (posters, abstracts, conference publications, and Wikipedia data) were excluded from the SLR. Other research articles concerned with the study area (global and national scale, MODIS imagery and products, local and urban scale, peat swamp, savannas, mangrove forests, forest fire, and wet/coastal forest studies, and studies on improving methodologies and estimation comparative studies) were also screened out from this SLR. While forest scale was not explicitly analyzed in this review, studies carried out at a local or hyper-local scale were not included. The published research articles with evaluation criteria (R²) were included in this review. The twenty-three (23) research studies having other accuracy criteria (Standard Deviation, RMSE) were included in this SLR. This resulted in the exclusion of 23 articles from the selected 98 articles during the meta-analysis of the accuracy assessment. The meta-analysis was performed only on the rest of the scrutinized research articles (75), which used a coefficient of determination (R²). The qualitative analysis was performed on all 98 selected articles, as shown in Figure 2.

2.5. Quality Assessment Criteria

The next step was to create the SLR quality assessment criteria. To ensure the quality and relevance of the scholarly literature involved in the review process, the abstracts of the articles were scrutinized to analyze and purify them during its implementation. These articles followed different methods and approaches for forest aboveground biomass estimation and mapping. Part of the meta-analysis assessment included all the criteria used to evaluate outcomes, including the coefficient of determination. The papers selected were relevant to estimating and mapping aboveground biomass in forests and fully met the quality assessment criteria explained above in Section 2.3 and Section 2.4. The reliability of the selected articles was improved by using inclusion and exclusion criteria in quality assessment.

Mendeley Desktop v1.19.8 software was maintained and assisted with all these steps. This software also reliably performed research article association, citation, and bibliography [38].

2.6. Segregation and Analytics

The scrutinized research studies were further categorized according to the use of spaceborne, airborne, or space- and airborne platform datasets for analysis. Country, study year, types of empirical models, remote sensing datasets, accuracy assessments, and forest biome insights were the key elements separated out for analysis. These analyses are explained in Section 3 of the Results and Section 4 of the Discussion to offer some insights. The majority of the selected publications used the coefficient of determination (R²), mean absolute error (MAE), normalized root mean square error (nRMSE), and root mean square error (RMSE) to evaluate the model accuracy. Of the above, the current review examined the coefficient of determination (R²) to measure the accuracy. This is because the evaluation by RMSE is based on the values of observation, which vary by type of forest and by region/country.

2.7. Scientometric Analysis

In this SLR, a scientometric investigation was also performed using CiteSpace Software 6.3.R1 Basic [32]. These analytics encompassed the webometrics, bibliometrics, and informetrics of the selected studies, further highlighting insights from the selected articles’ citations. An article’s citations and keyword connectivity were identified through this software. The node’s size showed the number of key phrases used in the literature analysis. The use of the above software has helped establish relationships between the publication’s titles and keywords, and has also aided in effective visualization.

3. Results

This is the most vital portion of this SLR. The results provide a comprehensive demonstration of the extracted information used in this study. The interpretation of data acquired from the article results and analysis is also described in the Discussion section of this SLR. The spatial distribution of the selected articles’ study areas is shown in Figure 3. Among a total of 98 selected studies, 54 were relevant to spaceborne platforms, 22 to airborne platforms, and 22 to space- and airborne platforms. Information regarding all articles (years, databases) is given in Figure 4 and Table 1.

Figure 3. Spatial distribution of selected article study areas.

Figure 4. Yearly distribution of the articles and their allocation in the three categories.

The maximum number of research articles (15) occurred in 2023 and 2019, followed by 2022, indicating a possible increase in research on forest aboveground biomass estimation and mapping in recent years, which will increase further with the launch of the ESA’s BIOMASS mission and the NASA–ISRO SAR mission (NISAR) and the continuity of GEDI datasets.

ScienceDirect was the most significant data source, with the maximum number of research publications (56), followed by MDPI (17) and Scopus (16).

The selected articles were classified into three categories (spaceborne, airborne, and space- and airborne) according to the taxonomical arrangement. Of the 98 articles, only 54 were relevant to spaceborne platforms, 22 were related to airborne platforms, and 22 were related to space- and airborne platform datasets.

3.1. Scientometric Analysis

CiteSpace software v.6.3.R1 is a powerful tool that offers citation analytics based on keywords, abstracts, and titles. Meanwhile, the meta-analysis followed the criteria defined in this review and the PRISMA framework guiding rules. Scientometric analysis offers detailed analytics related to citation catalogs like Web of Science (WoS) and Scopus. In this review, the articles were selected irrespective of the number ofcitations. Web of Science (WoS) provided 3244 citations to the scientometric scrutiny text list in CiteSpace Software for the selected articles the year-wise citation and publication trends as per the WoS analysis are shown in the graph in Figure 5. The citation trend line in the chart indicates that there was an increase in citations between the years 2015 and 2023. This further proves that scholars have actively shown interest in developing methods for mapping and estimating the aboveground biomass of forests using remotely sensed datasets in recent years. Thus, the 2015–2023 timeline selection was suitable, as evidenced by the citation trend graph.

Figure 5. Year-wise citation and publication trends (WoS).

A scientometric analysis for document co-citations of the scrutinized articles (98) was conducted using the default conditions with unlimited look-back years. In this study, the 292 merged network nodes for 1442 links show a link average amount of 4.93 per node. The outcomes of the co-citation clustering analysis are presented in Figure 6 and Figure 7, based on titles and keywords using CiteSpace. Among the clusters, ground biomass, optical satellite data, remote sensing data, Sentinel-2, and multi-sensor data synergy were the most relevant, followed by forest AGB, LiDAR, multi-sensor, laser scanning, and canopy height model. The critical theme of forest biomass estimation was also explored in the document co-citation analysis described in this review. This review used CiteSpace v.6.3.R1 to view the scholarly literature from 2015 to 2023. The analysis was performed on 20 May 2024, and it showed a network of 292 nodes and 925 edges with a density of 0.0218. Of all the nodes in the network, 1.0% were labeled, and the single most critical connected compound covered 100% of the network. The selection criteria consisted of LRF = −1.0 and G-index = 21. The pruned method used was Pathfinder, the weighted mean silhouette S = 0.8588, and modularity Q = 0.7364. The harmonic mean was found to be Q, S = 0.7929.

Figure 6. Co−citation analysis based on titles using CiteSpace. Numbers show title importance, and different colors show their cluster strength [32].

Figure 7. Co−citation analysis based on keywords using CiteSpace. Numbers show keywords’ importance, and different colors show their cluster strength.

3.2. Forest Aboveground Biomass Estimation

The various sensors, comprising spaceborne, airborne, and space- and airborne datasets, were integrated to estimate forest aboveground biomass, field inventory, and survey data. Therefore, this section is further divided into three categories (spaceborne, airborne, and space- and airborne datasets) for segregation and analytics information to explore the various approaches and results adopted in each dataset section.

3.2.1. Spaceborne Datasets

Using spaceborne data provides an excellent perspective for estimating and mapping forest AGB using various empirical regression models. The frequently utilized spaceborne datasets include optical, SAR, and spaceborne LiDAR. These datasets have the potential to provide more reliable information on land reflection in light of spatial, temporal, and radiometric advances. Moreover, these spaceborne datasets were integrated with field inventory and field plot data using various regression models and provided reliable information. After having classified the selected articles, this section includes studies (54) that describe the use of spaceborne datasets for forest AGB estimation and mapping. The graphs, as shown below in Figure 8, display the frequency distribution of the research articles in terms of datasets, regression models used, and article frequency per year and country/continent-wise.

Figure 8. Analysis of spaceborne dataset articles from 2015 to 2023. (A). Frequency distribution of articles country- and continent-wise. (B). Article frequency year-wise. (C). Year-wise frequency of regression models employed. (D). Year-wise frequency of spaceborne datasets employed.

These investigations were distributed geographically as follows: China (19), India (10), South Africa (4), Brazil (4), Italy (2), and Vietnam (2). Other countries were the USA, Pakistan, Japan, Mozambique, Madagascar, Iran, Germany, Gabon, Colombia, Canada, Zimbabwe, Tanzania, and Australia. These studies showed that spaceborne datasets, field data, regression models, and China (country) were unique characteristics pertinent to this integration approach. Most articles used the Sentinel (Sentinel-2 (6), Sentinel-1 and 2 (5)) and Landsat (Landsat-5, 7 and 8 (1), Landsat-7 and 8 (1), Landsat-5 (2), Landsat-8 (4)) imagery to estimate forest AGB, as shown in Figure 8D. Figure 9 shows a flowchart of spaceborne datasets, regression models used, and countries/continents.

Figure 9. Flowchart of the spaceborne category mentioning continents/countries, types of spaceborne datasets, and regression models employed.

The integration of different spaceborne datasets using regression models is essential not only to estimate forest AGB and its uncertainties but also to give information about which spaceborne datasets are valuable (coverage area, spatial and spectral resolution domain) to estimate biomass in a time- and cost-efficient manner. Therefore, a reliable selection of spaceborne datasets and regression models along with field data will not only aid in the estimation of forest AGB but also offer knowledge for utilization in decision-making for forest ecosystem management.

In this SLR, 54 papers described the utilization of spaceborne datasets to estimate forest AGB using regression models. Analysis of these selected articles (54) showed that twenty-three different regression models were used to integrate the spaceborne datasets with field inventory or field data. Among these selected articles (54), random forest (22) [16,17,26,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57], linear regression (6) [58,59,60,61,62,63], multi-linear regression (3) [64,65,66], stochastic gradient boosting (3) [67,68,69], and XGBoost (2) [10,70] were the most common models employed for the application of spaceborne datasets. Other regression models were used only once [71,72,73,74,75,76,77,78,79,80,81,82,83,84,85,86,87,88].

3.2.2. Airborne Datasets

The most frequently used airborne platform datasets mentioned in this category included airborne LiDAR, Airborne Laser Scanning (ALS), and UAV datasets. These datasets with high spatial resolution can provide improved forest AGB estimations over a local area. Moreover, these airborne datasets were integrated with field inventory and field data utilizing regression models to produce accurate and reliable information about the forest AGB. After the classification, the airborne datasets category comprised 22 articles featuring using airborne datasets. The graphs shown in Figure 10 display the frequency distribution of research articles in terms of airborne platform datasets, regression models, year-wise frequency, and country/continent-wise.

Figure 10. Visualization of airborne dataset articles from 2015 to 2023. (A). Frequency distribution of articles country- and continent-wise. (B). Article frequency year-wise. (C). Year-wise frequency of regression models employed. (D). Year-wise frequency of airborne datasets employed.

These investigations were distributed geographically as follows: USA (4), China (4), and French Guiana (3). Other countries were Tanzania, Switzerland, Spain, Portugal, Norway, Malaysia, Ghana, Gabon, Finland, the Czech Republic, and Brazil. These studies showed that airborne datasets, regression models, and the USA and China (countries) were unique characteristics pertinent to this integration approach. Most articles used the Airborne LiDAR (11) and ALS (3) datasets for the airborne datasets to estimate forest AGB, as shown in Figure 10D. Figure 11 shows a flowchart of airborne datasets, regression models used, and country/continents.

Figure 11. A flowchart of the airborne category mentioning the continents/countries, types of airborne datasets, and regression models employed.

Using regression models to integrate different airborne datasets is essential to estimate forest AGB and its uncertainties and to give information about which airborne datasets are valuable (coverage area and spatial and spectral resolution) to estimate biomass in a time- and cost-efficient manner. Therefore, a reliable selection of airborne datasets and regression models using field data will aid in the estimation of forest AGB. It shall also contribute towards technology utilization in improving estimation accuracies and reducing uncertainties.

In this SLR, 22 papers were found that described the utilization of airborne datasets to estimate forest AGB using regression models. In these selected articles (22), linear regression (2) [89,90], random forest (2) [91,92], and allometric equation (2) [24,93] models were employed the most. The other models were used once each in articles [94,95,96,97,98,99,100,101,102,103,104,105,106,107,108,109]. Linear regression and random forest were the most important models for the application of airborne datasets in forest AGB estimation.

3.2.3. Space- and Airborne Datasets

The most frequently used space- and airborne datasets mentioned in this category included airborne LiDAR, ALS, and UAV datasets, and optical, SAR, and spaceborne LiDAR datasets. These data integrations provide overall coverage of spatial, temporal, and radiometric variations within platform datasets to improve the estimations of forest AGB. Moreover, these space- and airborne datasets were integrated with field inventory/field data using various regression models and provided reliable information for the forest AGB estimation. After the classification, the space- and airborne datasets category comprised 22 articles describing the use of space- and airborne datasets in detail. In Figure 12, the graphs display the frequency distribution of research articles discussing spaceborne–airborne platform datasets, the regression models used, and country/continent-wise.

Figure 12. Visualization of space- and airborne dataset articles from 2015 to 2023. (A). Frequency distribution of articles country- and continent-wise. (B). Article frequency year-wise. (C). Year-wise frequency of regression models employed. (D). Year-wise frequency of space- and airborne datasets employed (AL = airborne LiDAR).

These investigations were distributed geographically as follows: USA (10), Japan (2), and China (1). Other countries were Thailand, Spain, Norway, Nepal, Mexico, Germany, Gabon, Canada, and Australia. These studies showed that space- and airborne datasets, regression models, and the USA (country) were unique characteristics pertinent to this integration approach. In the space- and airborne datasets category, most articles used airborne LiDAR, Landsat, Sentinel, and GEDI datasets to estimate forest AGB, as shown in Figure 12D. Figure 13 shows a flowchart of space- and airborne datasets, the regression models used, and the countries/continents.

Figure 13. A flowchart of space- and airborne category mentioning continents/countries, types of spaceborne–airborne datasets, and regression models employed.

Using regression models to integrate different space- and airborne datasets is essential to estimate forest AGB, reduce the uncertainties, and give information about which space- and airborne datasets are helpful with respect to area, resolution, and forest type.

In this SLR, 22 papers described using space- and airborne datasets to estimate forest AGB using regression models. Among these selected articles (22), eleven regression models were used to integrate space- and airborne datasets with the field inventory or field data. Among these selected articles (22), random forest (8) [110,111,112,113,114,115,116,117] and linear regression (5) [118,119,120,121,122] were the most often used models, and the other models were employed once [123,124,125,126,127,128,129,130,131].

3.3. Accuracy Assessment Analytics

A coefficient of determination (R²) is a statistical metric defining important insights about how a model works and how variance is distributed between independent and dependent variables. Analysis of the coefficients of determination (R²) is depicted as an error bar plot to obtain the performance information for the forest AGB estimation regression models. Each point represents the mean R² value for a study, while the horizontal lines (error bars) show the range from minimum to maximum R² values. These have been used to show how various research workflows relevant to remote sensing datasets and regression models performed in terms of consistency, reliability, and range of R². This section is also divided into three sections: (i) spaceborne analysis, (ii) airborne analysis, and (iii) space- and airborne analysis, organized in terms of the regression models and datasets used.

3.3.1. Spaceborne Analysis

Accuracy assessment for the spaceborne platform datasets was carried out for datasets, and a regression model on the article’s coefficient of determination (R²) was used. Forty-eight (48) articles [10,16,17,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,63,65,66,67,68,69,70,71,72,73,74,75,76,78,79,80,82,84,85,86,88] were employed for further analysis, as they had R² values among the total classified articles (54). The error bar plot of R² and the regression model used, as shown in Figure 14, showed that the random forest regression model (21) was the only model with a reasonable number for analysis, followed by linear regression (5). The highest number of studies with the random forest model were conducted in China (6). Domingues G. et al. [73] achieved an R² of 0.95 employing AVNIR-2 and PALSAR (as shown in Figure 14) datasets and an artificial neural network model with parameters (LVH, Green Band) for eucalyptus forest in Brazil, but lower in number. Wang Y. et al. [85] employed GLAS and field data with the PIPE method (plane inclined and point of signal ending) with parameters (terrain, slope) and achieved an R² of 0.92 for tropical rain forests in China. Guo Q. et al. [40] employed Sentinel-1 and 2 and GEDI datasets with the random forest model incorporating field plot data with parameters (FAPAR, EVI, CAB, CMH). They achieved an R² value of 0.89 for planted forests in Liaoning Province of China. The integration of Sentinel 1 and 2 and GEDI was a promising integration approach to improve the estimation but was lower in number in this review, as shown in Figure 15. Hlatshwayo S. et al. [52] employed Spot-6 imagery with a random forest model incorporating field data with parameters (homogeneity, entropy, mean, data range, correlation, variance, and Bands 1, 2, 3, and 4) and achieved an R² of 0.88 for tropical dry forests in South Africa. Dube T. et al. [69] employed Worldview-2 imagery (as shown in Figure 14) with a stochastic gradient boosting regression model with parameters (NDVI-RE, TVI.RE, IPVI.RE, GI.RE, SR, NDVIc, OSAVI, rainfall, temperature, total wetness index, insolation, and slope) and achieved an R² of 0.88 for eucalyptus and pine forest in South Africa. Mareya H. et al. [59] employed GeoEye imagery with a linear regression model and field data with parameters (canopy area). They achieved an R² of 0.87 for the deciduous miombo forest in Zimbabwe. Li H. et al. [44] employed PALSAR and Sentinel-2 datasets with a random forest model along with parameters (VH mean, VH variance, NDVI, B5, B9, B8a, B11, B12, SR, VH correlation) and showed the lowest accuracy with an R² of 0.37 for cypress forests in Japan.

Figure 14. Graph showing spaceborne regression model-based accuracy assessment analysis (the orange box represents the mean R² value, and the horizontal lines show the minimum to maximum R² values).

Figure 15. Graph showing spaceborne dataset-based accuracy assessment analysis (the orange box represents the mean R² value, and the horizontal lines show the minimum to maximum R² values).

The error bar plot of R² and the spaceborne datasets used is shown in Figure 15. It shows that Sentinel (10) was the most utilized imagery in studies with a reasonable number for analysis (Sentinel-2 (6), Sentinel-1 and 2 (4)), followed by Landsat (8) (Landsat-5 (3), Landsat-8 (3), Landsat-5, 7, and 8 (1), and Landsat-7 and 8 (1)). The highest number of studies utilizing Sentinel imagery and integrating it with other datasets was in China (14). Although not well explored, another promising integration framework for the spaceborne datasets of AVNIR, PALSAR, GLAS, Sentinel-1 and 2, and Spaceborne LiDAR (GEDI) resulted in an R² of (0.95, 0.92. 0.89) [40,73,85].

3.3.2. Airborne Analysis

An accuracy assessment of the airborne platform datasets and a regression model grounded on the article’s coefficient of determination (R²) was carried out for the datasets. Thirteen articles [91,93,94,95,96,97,98,99,100,102,103,105,107] were employed for further analysis, as they had R² values among the total classified articles (22). The error bar plot of R² and the regression models used in Figure 16 showed that stepwise regression was the only model with two articles for analysis. The highest numbers of studies were conducted in the USA (4) and China (4).

Figure 16. Graph showing airborne regression model-based accuracy assessment analysis (the orange box represents the mean R² value, and the horizontal lines show the minimum to maximum R² values).

Ni-Meister W. et al. [97] employed LVIS and airborne LiDAR datasets with geometric optical and canopy radiative models with parameters (tree height, stem diameter, vegetation density, leaf and background reflectivity coefficient ratio, biomass indices). They achieved an R² of 0.89 for temperate deciduous and conifer forests in the USA. Liu X. et al. [107] employed airborne LiDAR using an empirical logarithmic model with parameters (covariance matrix) and achieved an R² of 0.88 for tropical forests in Gabon. Meng S. et al. [102] employed airborne LiDAR and aerial imagery data with a support vector regression model along with parameters (wavelength, pulse length, sampling interval, sensor size, pulse firing rate, sensor size, and dimensions, r-spectrum, wavenumber). They achieved an R² of 0.88 for broadleaf and conifer forests in China. Cao L. et al. [99] achieved an R² of 0.87 employing airborne LiDAR data using a multiple regression model with parameters (percentile heights h25–h95, height variation) for Moso Bamboo Forest in China. Ramachandran N. et al. [105] utilized a quadratic model to achieve an R² of 0.86 employing airborne SAR and LiDAR datasets with parameters (backscattered power, FHandQ, HV) for moist tropical forests in French Guiana. Torre-Tojal L. et al. [91] achieved an R² of 0.7 employing airborne LiDAR using a random forest model with parameters (height metrics (Elev P95, P99, P90)) for pine forests in Spain. Liao Z. et al. [95] achieved an R² of 0.7 employing airborne SAR and airborne LiDAR datasets with the RVoG model and parameters (volume backscatter from the forest canopy) for tropical rainforest in French Guiana. De Almeida C. et al. [96] achieved an R² of 0.70 employing airborne LiDAR and hyperspectral data using a linear model with ridge regularization and parameters (upper canopy cover and height percentiles, NIR, SWIR) for the Amazon Forest in Brazil. The error bar plot of R² and the airborne platform datasets used is shown in Figure 17. Airborne LiDAR was the most utilized dataset in studies with a reasonable number (6) for analysis. Most airborne and UAV LiDAR studies were conducted in China (2). Another promising framework was integrating airborne LiDAR and aerial imagery [102].

Figure 17. Graph showing airborne datasets-based accuracy assessment analysis (the orange box represents the mean R² value, and the horizontal lines show the minimum to maximum R² values).

3.3.3. Space- and Airborne Analysis

An accuracy assessment of the space- and airborne platform datasets was carried out, and a regression model grounded on the article’s coefficient of determination (R²) was used. Fourteen articles [111,112,113,114,116,118,119,120,121,124,125,126,127,128] were employed for further analysis, as they had R² values among the total classified articles (22). The error bar plot of R² and the regression models used, shown in Figure 18, showed that random forest was the most used model, with five articles for analysis, followed by linear regression (4). The highest numbers of studies were conducted in the USA (7) and China (1).

Figure 18. Graph showing spaceborne–airborne regression model-based accuracy assessment analysis (the orange box represents the mean R² value, and the horizontal lines show the minimum to maximum R² values).

Qi W. et al. [128] employed GEDI, TDX, and ALS datasets with multiple linear models with parameters (TDX height) and achieved an R² of 0.87 for mixed temperate forests in the USA. Cooper S. et al. [125] employed Landsat-7 and AVIRIS datasets with Gaussian process regression with parameters (combined EnMAP and Landsat texture signatures). They achieved an R² of 0.87 for conifer and deciduous forests in the USA. Zhang Z. et al. [124] employed Landsat-5, PRISM, and UAVSAR datasets with multivariate regression models with parameters (backscatter from track and canopy height). They achieved an R² of 0.85 for boreal and temperate forests in the USA. Narine L. et al. [120] employed ALS and ICESat-2 datasets with linear regression models with parameters (mean height, height bin, PCL-derived 10th percentiles, and canopy cover). They achieved an R² of 0.79 for pine forests in the USA. Zald H. et al. [111] employed PRISM and PALSAR imagery with a random forest model with parameters (NDVI, NDWI, Chlorophyll-VI). They achieved an R² of 0.69 for pine and coniferous forests in Germany. Woodcock C. et al. [112] employed Sentinel-1, Landsat-8, and NAIP datasets with random forest models with parameters (NDVI, EVI, VV, VH, brightness, greenness, wetness, heights 95, 85, 75, 50 percentile, image texture, texture ratio). They achieved an R² of 0.63 for pine forest in the USA. Trung H. Nguyen et al. [114] employed Landsat-TM/ETM and airborne LiDAR datasets with random forest with parameters (NBR, TCA, TCG, TCW, elevation, slope, stem density, basal area) and achieved a lower R² of 0.59 and mean R² of 0.56 for the boreal coniferous forest in Australia.

The error bar plot of R² and the space- and airborne datasets used showed that Landsat-5, 7, and 8 and airborne LiDAR (3) were the most utilized imagery in studies for analysis, followed by GEDI, TDX, ICESat-2, and ALS with the lowest numbers (2), as shown in Figure 19. The highest number of studies with airborne LiDAR, ALS, Sentinel, GEDI, ICESat-2, TanDEM-X, and Landsat datasets were conducted in the USA (7). Another promising integration framework employed LiDAR (GEDI), ALX, and TDX and achieved an R² of 0.87 [128].

Figure 19. Graph showing space- and airborne-based accuracy assessment analysis (the orange box represents the mean R² value, and the horizontal lines show the minimum to maximum R² values).

Summarized information compiled regarding the area, data source, research highlights, and accuracy of the main articles in all three categories (with respect to the highest coefficient of determination (R²) in that category) is shown in Table 2. It provides details on the most commonly used data sources, along with the parameters used and their accuracy in one location. It also serves to highlight the effectiveness of multi-platform integration approaches in forest AGB estimation, which was the main objective of this SLR. Table 2 also provides useful insights and is aligned with the error bar plots, which provide information regarding datasets and regression models employed, as shown in Figure 14, Figure 15, Figure 16, Figure 17, Figure 18 and Figure 19.

Table 2. Summarized information on the main articles in all three categories.

Study Area	Data Source	Research Highlights	Accuracy	Reference
Spaceborne Dataset Articles
Minas Gerais State, Brazil	AVNIR-2, PALSAR	An artificial neural network with parameters (NIR, B, G, R, polarization levels) was employed.	R2 = 0.95	[73]
Yunnan, China	GLAS	The plane inclined and point of signal ending (PIPE) method was used with parameters (terrain, slope).	R2 = 0.92	[85]
Liaoning Province, China	Senitinel-1 and 2, GEDI-L2A	The random forest model was employed with parameters (FAPAR, vegetation cover, CAB, CWC, vegetation index).	R2 = 0.89	[40]
KwaZulu-Natal, South Africa	Spot-6	The study employed a random forest model with parameters (homogeneity, entropy, mean, data range, correlation, variance, and Bands 1, 2, 3, and 4).	R2 = 0.88	[52]
Harare, Zimbabwe	GeoEye-1	The linear regression model was employed with parameters (canopy area).	R2 = 0.87	[59]
Airborne Dataset Articles
Sierra Nevada, USA	LVIS Airborne LiDAR	The geometric optical and canopy radiative model was employed with parameters (tree height, stem diameter, vegetation density, leaf and background reflectivity coefficient ratio, and biomass indices).	R2 = 0.89	[97]
Lope, Gabon	Airborne SAR, Airborne LiDAR	An empirical logarithmic model with parameters (covariance matrix) was employed.	R2 = 0.88	[107]
Heilongjiang, China	Airborne LiDAR, Aerial Imagery	Support vector regression was employed with parameters (wavelength, pulse length, sampling interval, sensor size, pulse firing rate, sensor size, dimensions, r-spectrum, and wavenumber).	R2 = 0.88	[102]
Jiangsu Province, China	Airborne LiDAR	The multiple regression model was employed with parameters (percentile heights h25–h95, variation of height).	R2 = 0.87	[99]
Paracou, French Guiana	Airborne LiDAR, Airborne SAR	The study employed a quadratic model with parameters (backscattered power, FHandQ, HV).	R2 = 0.86	[105]
Spaceborne and Airborne Dataset Articles
New Hampshire, USA	ALS, GEDI, TanDEM-X	The multiple regression model was employed with parameters (TDX height).	R2 = 0.87	[128]
California, USA	Landsat-7, AVIRIS, EnMap Imagery, Airborne LiDAR	The Gaussian process regression model was used with parameters (EnMAP and Landsat spectral metrics).	R2 = 0.87	[125]
Howland, USA	Landsat-5, PRISM, UAVSAR	The multivariate regression model employed parameters (backscatter from track and canopy height).	R2 = 0.85	[124]
Texas, USA	ALS, ICESat-2	The linear regression model with parameters (mean height, height bin, PCL-derived 10th percentiles, and canopy cover) was employed.	R2 = 0.79	[120]
Hokkaido, Japan	Airborne LiDAR, TanDEM-X	The random forest model was employed with parameters (annual precipitation, TDX-CHM, forest type, HV).	R2 = 0.70	[113]

3.4. Remote Sensing Platform Datasets in Forest Biome

The information extracted from the selected articles included forest types, survey data, and forest biome information. The forest types were further grouped by forest biome to understand their relation with remote sensing datasets. The forest biome vocabulary was drawn from the article [132], which provides useful information about utilizing remote sensing datasets concerning the forest biome, as shown in Figure 20, Figure 21 and Figure 22. It gives an overview of platform-wise dataset segregation within the forest biome. It provides insights for researchers to decide which kinds of datasets would be better to select for accurately estimating forest AGB biomass, along with machine learning models and field data concerning forest biome. The analysis showed that most studies carried out utilizing spaceborne datasets were most frequently conducted for temperate broadleaf and mixed forests (9), airborne datasets were used most for tropical rain forests (5), and space- and airborne datasets were used most for temperate coniferous forests (9). Information regarding the reference, study area, survey data, forest type, and forest biome for each of the selected articles (98) is attached as Appendix A, Appendix B and Appendix C.

Figure 20. Utilization of spaceborne platform datasets by forest biome.

Figure 21. Utilization of airborne platform datasets by forest biome.

Figure 22. Utilization of space- and airborne platform datasets by forest biome.

4. Discussion

4.1. Discussion

Estimating forest aboveground biomass (AGB) with multiple remote sensing datasets is crucial for determining carbon density and stock, supporting and contributing to global carbon neutrality, and sustainable forest management initiatives. Previous studies have attempted to review various remote sensing datasets and integration workflows for forest biomass estimation [1,8,14,31,32]. However, these reviews have yet to synthesize platform-wise synergetic integration approaches to map and estimate forest AGB.

This paper reviews the outcomes and highlights the novelty and effectiveness of various approaches using diverse remote sensing platform datasets, machine learning models, and parameters for forest AGB mapping and estimation. This SLR categorized multi-source remote sensing datasets into three categories and revealed that every category efficiently contributes to forest AGB estimation using field data and machine learning regression models. This review analysis highlighted the most utilized datasets (Sentinel, Landsat, airborne LiDAR, and GEDI) among the remote sensing datasets and their integration approaches. Many machine learning approaches were utilized in the selected articles, but that random forest was the preferred regression model was evident. It was also linked with co-citation analysis regarding the selected articles’ keywords and titles, which supports the reliability of our review criteria process and validates the significance of our research objectives. This review also identified important parameters across various remote sensing datasets, further strengthening the novelty of our approach and highlighting the multi-dimensional nature of AGB estimation. Integrating AVNIR-2 and the PALSAR dataset using an artificial neural network shows the innovative techniques employed to reduce validation errors in forest estimation, yielding a virtuous R² value of 0.95 [75]. Similarly, Sentinel-1 and 2, spaceborne LiDAR (GEDI), TanDEM-X, and InSAR, in combination with a random forest model and TDX height model, demonstrate the efficacy of spaceborne datasets and ensemble learning approaches in achieving accurate AGB estimations [42,130]. This review also highlighted information about the various remote sensing platform datasets selected for forest biomes and revealed which datasets produce good results. Spaceborne instruments were primarily utilized in studies because of their wide availability and their ability to capture large-scale changes over time. They also offer a cost-efficient solution for continuous monitoring of forest environments. Another factor contributing to the frequency of spaceborne studies is the technological advances in satellite sensors in the last decade, such as Sentinel-2, Landsat-9, and GEDI, which have significantly enhanced their spatial, spectral, and structural information for more detailed and accurate forest AGB estimation. These are promising remote sensing datasets to integrate, as they help improve the efficiency of AGB modeling approaches and develop new approaches to reduce uncertainties and errors in forest biomass estimation.

4.2. Uncertainty in Forest AGB Estimation

To date, extensive global research has been conducted on increasingly reliable forest aboveground biomass estimation methods using remote sensing datasets and their integration approaches [44,95,128,133]. The accuracy of those approaches has been a reason for concern because these assessments must be carried out precisely to combat global warming and reach the global carbon emission neutrality targets effectively [133]. The accuracy of those methodologies has been something to frown upon. Several potential errors may overestimate or underestimate the AGB levels in the forest. Generally, stand structure, machine learning models, sample plot survey data, and remote sensing datasets add to these uncertainties [46,60,76,107]. However, the Sentinel-1A C-band cannot aid as an active predictor due to optical saturation error due to dense canopies and relatively lower penetration ability [77]. GEDI requires adjustment for multifaceted topography to confirm alignment with ground location, and due to its geolocation error, it overestimates forest AGB [44,61]. Using a random forest model does not include spatial patterns during modeling and may lead to spatial mismatch errors [48]. Many other uncertainties within the datasets may be attributed to sensor spatial resolution, sampling error, errors in allometric equations, and prediction model-associated errors [80]. These errors have resulted in uncertainties after the upscaling of field plot data and the unavailability of site-specific allometric equations in the estimated AGB product [53]. The errors of height variations in estimating forest AGB can result from the different growing stages of a single species [100]. Sunlight may also be a significant factor in identifying species configuration and extracting tree height and DBH using a LiDAR sensor in a dense forest canopy [110]. The LiDAR-derived AGB is uncertain when used as reference data instead of LiDAR field or plot data [119]. Quantifying and reducing uncertainties in the estimation remains a significant challenge. Seeing the diverse sources of uncertainties, it is essential to comprehensively compute the errors to reduce their effect on estimation accuracy. This offers a meaningful way forward for forthcoming research on reducing uncertainties and improving estimation accuracy. Furthermore, the accuracy of the estimation of forest AGB could be enhanced by using remote sensing dataset integration approaches, customized machine learning models, parameters, and field data insights.

4.3. Limitations

This systematic literature review provides a valuable overview of integration approaches combining remote sensing datasets with regression models for forest aboveground biomass estimation. Notably, this review features some limitations that can be addressed in future research. The scope of the present study also does not include analyses of forest characteristics (structures, species compositions, homogeneity, or heterogeneity). As stated above, analyzing forest characteristics may allow for tailored conservation strategies, promoting biodiversity and ecosystem resilience. Similarly, the scale of forest coverage has also not been analyzed in the present study. Changes in scale may increase or subsume complexity and detail and, therefore, determine the tools and techniques used for AGB estimation. Furthermore, the present study does not provide a detailed insight into forest AGB uncertainties and errors. Further detailed analysis of the uncertainties of data acquisition techniques and AGB modeling would be of interest for further refining remote sensing techniques, calibration models, and statistical approaches. This, in turn, would help understand and reduce the uncertainty in carbon stock assessments essential for global climate change mitigation initiatives.

4.4. Future Research Scope and Recommendations

Future research directions are suggested, along with knowledge gaps, for researchers from the analysis of this SLR based on the use of integrated remote sensing datasets to expand the scope and accuracy of forest AGB estimation. During this SLR, it was identified that there needs to be more consistent research to highlight the remote sensing dataset’s accuracy. Repeated research with integrated datasets using various regression models with different study areas will help improve the accuracy of AGB estimation. Integrating more parameters (satellite-derived independent variables, field plots, field inventory) will help link various integrated remote sensing datasets. This information will help to improve model validation and scope for reliable estimation workflows. Ground data validation should be the crucial approach in spaceborne LiDAR (GEDI, ICESat). With the help of ground validation, the uncertainty equation development should respond and help reduce the overall uncertainties in forest AGB estimates. Integrating frameworks with field inventory and plot datasets will provide different AGB estimates for study areas, forest types, and regions. These frameworks should also help capture forest changes over time but are crucial in enhancing machine learning model accuracies for AGB estimations. Technological advances like machine and deep learning can improve regression models. This, however, is a requirement to retrieve datasets and ground information for model training.

Additionally, future studies must prioritize the development of uncertainty equations for spaceborne LiDAR datasets like GEDI, which is crucial for refining AGB estimation approaches. In the future, initiatives such as the ESA’s BIOMASS mission, NASA–ISRO SAR (NISAR) missions, and the continuity of GEDI datasets hold the potential for enriching remote sensing datasets and emphasizing the course toward more effective monitoring and aligning with global objectives such as carbon neutrality under initiatives like REDD+ and the Paris Agreement.

5. Conclusions

This systematic literature review provides a comprehensive review of forest AGB estimation and shows the efficiency of integrating spaceborne and airborne datasets that align with the objectives. It can be concluded from the results of the SLR that integrating these datasets emerges as a dynamic strategy to improve biomass estimation and mapping, combining the diverse capabilities of spaceborne and airborne platforms. Considering the heterogeneity of forest environments, it is difficult to declare one specific remote sensing dataset and regression model as the best among others. However, methodologies like the random forest regression model demonstrated significant utility, particularly in regions like China. The limited number of studies in this field indicates a research gap, needing further investigation into synergistic integration approaches to enhance accuracy and reliability in AGB estimation. The approaches based on integrating Sentinel, Landsat, GEDI, and Airborne LiDAR datasets, along with a random forest regression model, may be strongly recommended to address this gap. These approaches mitigate uncertainties and support informed decision-making in forest ecosystem management.

Uncertainty in forest AGB estimation remains a critical challenge due to potential errors in remote sensing datasets, sample plot data, geo-location inaccuracies, and machine learning models. Addressing these uncertainties through integrated datasets and advanced modeling techniques is essential for improving estimation accuracy and supporting global climate goals. Future research should prioritize integrating diverse remote sensing datasets with robust ground validation to enhance AGB estimates further. Leveraging the advances in machine learning and developing uncertainty equations for spaceborne LiDAR, along with initiatives like the ESA’s BIOMASS and the NASA–ISRO SAR missions, is crucial. These steps will enrich data quality, align with global environmental goals, and contribute to sustainable forest management practices.