1. Introduction
Human encroachment and intensified agricultural activities have threatened the function, and even existence, of tropical peat swamp forests in Southeast Asia [
1,
2,
3]. The restoration of these degraded tropical peat swamp ecosystems poses an enormous global challenge for both scientists and policymakers alike [
4,
5,
6]. Tropical peatland restoration is supported by a number of both national and international goals relating to ecosystem management, climate change mitigation, and sustainable development. For example, peatland restoration is recognised as a key focus in reaching carbon emission reduction targets under the Kyoto Protocol [
7] and as a terrestrial ecosystem management and protection aim under the United Nations Sustainable Development Goal number 15 (SDG 15). Therefore, there is an urgent need for the development of landscape-scale measures for verifying and quantifying restoration procedures in order to achieve the goals and commitments related to tropical peat swamp forest recovery.
Mapping aboveground biomass (AGB) dynamics and assessing their changes in response to anthropogenic activities is critical to understanding and monitoring the current status of tropical peatlands, as well as the performance of restoration practices [
8,
9,
10,
11]. The collection of AGB data has traditionally been based on forest inventory plots [
12]. However, despite generating accurate results [
13], these ground-based methods have been deemed expensive, time-consuming, and often limited by their sample size and spatial distribution over landscape scales [
14,
15]. Advances in remote sensing technology and improved computing and image processing techniques have made it possible to estimate large-scale forest AGB in a robust, cost-effective manner [
16,
17,
18]. In particular, spaceborne sensors have generated a vast archive of data at a spatial and temporal scale appropriate to support ecosystem restoration and monitoring programs [
19,
20]. For example, the European Space Agency’s Sentinel-2 constellation, which carries the MSI sensor with improved spatial and temporal resolution over existing optical spaceborne sensors (e.g., the Landsat series) and unique spectral band selection, promises to benefit global forest AGB monitoring efforts [
21,
22]. Optical sensors, such as those aboard Sentinel-2, are often limited to recording the reflectance in the topmost part of a forest canopy. The restrictions of optical data in delivering detailed vertical information on forest stand structure have directly impacted the accuracy of forest AGB assessment [
23]. Despite this shortcoming, spaceborne optical imagery remains a popular data option in large-scale forestry applications due to the wide availability of data sources, access to archival datasets, and low cost (at times with no acquisition cost).
Aside from optical remote sensors, AGB information has been derived from active sensors such as synthetic aperture radar (SAR) and light detection and ranging (LiDAR), on space-borne and airborne platforms respectively [
24,
25]. Indeed, LiDAR has become one of the most prominent remote sensing technologies in forest assessment [
26,
27,
28,
29]. LiDAR metrics, from both discrete return (DR) and full waveform (FW), have been shown to obtain highly accurate forest AGB and above carbon density (ACD) estimates, with no saturation issues associated with areas of high AGB [
30,
31]. This lack of saturation is a key benefit associated with the use of LiDAR data in tropical forest ecosystems, where traditional optical data may struggle to provide accurate estimates in areas with dense vegetation cover due to signal saturation in bands associated with high reflectance in green vegetation (such as green and near-infrared bands). Further, LiDAR structural information, in particular from LiDAR metrics related to the midcanopy forest structure, is important for assessing biomass in tropical peat swamp forest AGB estimation [
9]. Despite these successes [
32], LiDAR surveys over landscape scales are expensive and problematic. LiDAR systems are often limited by their spatial and temporal coverage (airborne acquisition), thus producing results best suited to sample transect datasets rather than the wall-to-wall coverage often called for in regional forestry applications, particularly when the demand is for high spatial resolutions [
33,
34]. To combat these limitations, many studies combine LiDAR and spaceborne optical data for landscape-scale forestry applications [
35,
36,
37,
38,
39]. Here, LiDAR-derived forest attributes or models are commonly adopted as surrogate ground truth data. The large amount of training samples generated from LiDAR data can significantly improve the performance and accuracy in predicting forest attributes using machine learning algorithms [
40,
41], thus overcoming the restrictions posed by traditional field sampling methods.
This research focuses on the North Selangor Peat Swamp Forest (NSPSF), a large tropical peat swamp forest in Peninsular Malaysia under varying degrees of both growth and degradation from multiple drivers. In view of the growing international interest in tropical peatland conservation and restoration, NSPSF was under state protection with an established site-specific Integrated Management Plan (IMP) [
42] set out from 2014 to 2023 to support and implement tropical peat swamp restoration. The IMP specifically targeted key areas of rehabilitation, fire prevention, and buffer zone management. To date, several IMP interventions have been implemented in NSPSF with varying degrees of success, including drain blocking, fire prevention, and forest replanting programmes. The years 2015–2018 marked a period of significant effort in the initial trial and installation of several peat swamp management and restoration activities set out by the IMP. Evidence of the impacts (both positive and negative) of these activities on the restoration of NSPSF is required to aid local stakeholders in prioritising and formulating future management strategies. In this research, we test and demonstrate the use of Sentinel-2 satellite data in delivering accurate, high-resolution AGB maps of NSPSF for the years 2015 and 2018 using a Random Forest machine learning method. Despite the proven accuracy of LiDAR data in tropical peat swamp forest AGB estimation [
9] and the limitations associated with forest stand structural representation in optical data, Sentinel-2 was selected as the practical data option for this research based on data cost, coverage, and cloud contamination. A reference dataset derived from the LiDAR AGB model results by Brown et al. [
9] was used as training and validation data in the Random Forest model. The resulting AGB maps will facilitate a change detection analysis to assess AGB dynamics between 2015 and 2018, providing insights into the effectiveness of peat swamp management activities in NSPSF. This evaluation is crucial for informing future management strategies and supporting international conservation goals.
2. Materials and Methods
2.1. Study Site
NSPSF is located on the flat coastal plains of northwest Selangor, Peninsular Malaysia (3°33′31N, 101°18′21E). Comprising four forest reserves—Raja Musa Forest Reserve, Sungai Karang Forest Reserve, Bukit Belata Extension Forest Reserve (partial), and Sungai Dusun Forest Reserve—NSPSF is one of the largest remaining peat swamp forests in Peninsular Malaysia [
43], covering a total area of 81,304 ha. NSPSF provides a wealth of ecosystem services—such as habitat biodiversity, carbon sequestration, regulation of freshwater flow and flood mitigation, and climate regulation—that are all essential to the local livelihood of neighbouring human communities [
42]. However, decades of peat swamp forest degradation and encroachment have restricted NSPSF ecosystem function and the services it can provide [
44,
45,
46].
The forest is a recovering secondary mixed swamp forest in which multiple cycles of selective logging have created a complex, multi-layered canopy structure, crisscrossed by a 500 km network of log extraction canals [
42,
47]. The construction of the canals has scarred the forest, accelerating the drainage of the peat soil and consequently increasing fire risk for the reserve and surrounding area. The land use adjacent to the reserve is dominated by large-scale oil palm plantations. Conversion to industrial oil palm has transformed the landscape in North Selangor [
48], with reserve encroachment and hydrological consequences of plantation drainage posing a serious threat to the protection and ecosystem function of NSPSF. The reserve also notably borders the Tanjung Karang Irrigation Scheme to the west, which predominantly serves the Sekinchan rice paddy fields and mining/ex-mining activities to the southeast.
As a result of destructive human activities and the degradation of tropical peat swamp forests, the Selangor State Government agreed on a 25-year moratorium on timber extraction across the State starting in 2007 [
42]. Following this step toward peatland restoration and rehabilitation, the IMP was developed in 2013–2014, prepared by the Selangor State Forestry Department with technical support from the Global Environment Centre (GEC), and approved by the Selangor State Government in October 2014. This plan was a significantly updated version of a previous IMP for NSPSF incorporating research findings from within the reserve and best management practices for peat swamp conservation. The IMP specifically targeted key areas for rehabilitation, fire prevention, and buffer zone management across NSPSF, adopting a zonal approach where focused activities of restoration and conservation could be strategically directed to the areas/zones most at risk of degradation. Several IMP interventions were implemented in NSPSF with varying degrees of success, including drain blocking, the introduction of strict no-burning requirements, community-based fire prevention and patrolling programmes, and forest replanting in severely degraded sites.
2.2. Workflow
Figure 1 illustrates the workflow for estimating tropical peat swamp forest AGB across the entire NSPSF and assessing the impact of natural and anthropogenic change on the site between 2015 and 2018. This includes the extraction of Sentinel-2 model input variables (spectral and textural), the development of a predictive model for tropical peat swamp forest AGB, and a change analysis of 2015 and 2018 AGB map outputs.
2.3. Sentinel-2 Data and Preprocessing
Sentinel-2 is an Earth observation system part of ESA’s Copernicus programme—a series of satellites under the family name Sentinel with a policy of free and open-access spaceborne data. Sentinel-2 is a constellation of twin satellites, Sentinel-2A and Sentinel-2B. The pair aims to provide enhanced multispectral data with a high revisit time of 5 days [
49]. The Sentinel-2 satellites carry the MSI sensor that samples across thirteen spectral bands at a spatial resolution ranging from 10 to 60 m. The MSI sensor’s unique band combination importantly includes three vegetation-focused bands positioned in the Red-Edge zone (across 705–865 nm, at a 20 m spatial resolution).
Sentinel-2 MSI scenes were acquired for 31 December 2015 and 13 February 2018 (both dates fall in the region’s dry season), supplied as Level 1C geocoded formatting. As Sentinel-2 Level 1C data are top-of-atmosphere reflectance, the ENVI FLAASH algorithm was applied to convert this to surface reflectance (bottom of atmosphere). MSI spectral bands with a 20 m spatial resolution (band 5, band 6, band 7, band 8a, band 11, and band 12) were resampled to 10 m by applying the nearest-neighbour interpolation algorithm. Both Sentinel-2 MSI scenes were projected to WGS 1984 UTM Zone 47N (the same projection as Brown et al. [
9] NERC Airborne Research Facility LiDAR data), and key structures in the landscape (using GPS data collected in the field) were used to check the intercalibration of the two scenes.
2.4. Sentinel-2 Model Input Variables
A total of 142 spectral and textural variables were calculated to estimate the AGB of NSPSF for the two dates of interest. Spectral variables include the preprocessed Sentinel-2 MSI bands (band 2: blue (490 nm), band 3: green (560 nm), band 4: red (665 nm), band 5: vegetation Red-Edge (705 nm), band 6: vegetation Red-Edge (740 nm), band 7: vegetation Red-Edge (783 nm), band 8: near-infrared (842 nm), band 8a: narrow near-infrared (865 nm), band 11: short wave infrared (1610 nm), and band 12: short wave infrared (2190 nm)) and a selection of vegetation indices (
Table 1). The selection of vegetation indices was based on their performance in past research [
50,
51]; these indices have proven to represent the strong contrast in reflectance in the near infrared (NIR) by plant matter and the strong absorption by chlorophyll in the red part of the electromagnetic spectrum, essentially measuring vegetation greenness. Additionally, vegetation indices adapted to utilise the Red-Edge zone were included to take full advantage of the multiple bands covering this sensitive spectral region [
52,
53]. Vegetation indices were calculated in R software (version 3.5.2) using the ‘Raster’ package [
54].
Texture variables based on the grey-level co-occurrence matrix (GLCM) approach [
55] were generated to enhance the spatial discrimination of features in the data scenes, such as tree density and linear structures in the landscape (i.e., canals). Three GLCM metrics—homogeneity, mean, and variance (
Table 2)—were applied to individual Sentinel-2 MSI bands under different windows sizes (3 × 3, 5 × 5, 7 × 7, and 9 × 9 pixels) at an offset distance of 1 and averaged in all directions. Texture variables were generated in R software (version 3.5.2) using the ‘glcm’ package [
56].
2.5. Training and Validation Data
2.5.1. ‘LiDAR AGB Model’
In order to overcome the many barriers to obtaining field inventory plot data for AGB estimation, past studies have adopted LiDAR transects as surrogate plot data [
35,
36,
37,
38]. In a similar vein, this research uses the site-specific model developed by Brown et al. [
9] as model training and testing data for predicting AGB for NSPSF. The predictive model and statistics presented are listed in
Table 3. Based on this model, a thematic map of tropical peat swamp forest AGB was generated for a subset of LiDAR flightlines recorded by the NERC ARF team in the 2014 Malaysia campaign, gridded at 10 × 10 m spatial resolution (scale matching the Sentinel-2 MSI 10 m spatial resolution bands). It is important to note here the three-year time delay between the collection of the airborne LiDAR data (2014) and field plot inventory data (2017) used by Brown et al. [
9] in developing this model. This may introduce error into the AGB model; however, all field plots were established to be free from logging and fire, with no obvious disturbance within the three years.
The AGB data generated from the Brown et al. [
9] model, herein known as the ‘LiDAR AGB model’, represent an area of 9457.08 ha with high accuracy, covering just over 10% of NSPSF (
Figure 2). These sections of the reserve have faced little reported large-scale forest cover change since the updated IMP implementation in 2014. The areas of the reserve covered by the ‘LiDAR AGB model’ support forest with a range of structural characteristics (i.e., height and density), and are thus deemed appropriate and representative as training and testing data for mapping AGB across the whole of the NSPSF (
Figure 2). A simple visual check on the spatial positioning of datasets was made. Sourcing appropriate structures in the landscape as reference points was restricted due to the forest structure and data resolution (10 m), with reference points limited to clear reserve boundary lines. The spatial correspondence of datasets in this research may introduce errors into the final wall-to-wall AGB estimations.
The ‘LiDAR AGB model’ totaled 949,119 pixels (10 m spatial resolution), of which 70% (664,383) were randomly sampled to use as training data, and the remaining 30% (284,736) were reserved for testing (
Table 4).
2.5.2. Field Inventory Data
Forest inventory ground truth data were collected in 2018 to be used as validation data for the 2018 AGB model output. To estimate field AGB, 17 circular field plots with a radius of 15 m were established throughout the reserve, with a keen effort to cover areas of different locations, forest types, degradations, and management in NSPSF.
The location of the central tree, and thus the centre point of the plot, was recorded by a field GPS (Garmin GPSMAP 64), with an average horizontal error of <5 m when measurements were averaged over ten minutes. The positions of all further trees within the plot were noted relative to the centre (distance and angle). For each plot, all live trees with a diameter at breast height (DBH) of >5 cm were measured, and recognised tree species were recorded. Additionally, dependent on the top of tree visibility, selected trees of varying height classes were measured by laser height meter. Plot AGB values were calculated based on the general pan-tropical allometric equation developed by Chave et al. [
58].
where ρ = wood density (g/cm
3), D = DBH (cm) and H = height (m).
Inputs included field-measured DBH, species (to estimate wood density), and tree height when available (missing height data were retrieved from the Feldpuash et al. [
59] averaged model, with the region set to Southeastern Asia). Wood density was estimated for the trees based on their taxonomy using the global wood density database (GWD) [
60,
61], unidentified trees or trees with a missing genus in the GWD were assigned a mean plot-level wood density value.
The ‘BIOMASS’ package in R software (version 3.5.2) [
62] was used to run AGB estimation. AGB values were calculated for individual trees within each 15 m radius field plot and then summed to obtain total AGB plot values.
2.6. Random Forest Regression
The Sentinel-2 2015 spectral and textural variables and the training data (70% of the ‘LiDAR AGB model’) were used as inputs for a nonparametric data fusion machine learning algorithm, Random Forest [
63]. As previously mentioned, the ‘LiDAR AGB model’ was developed from a LiDAR dataset captured in 2014 and forest inventory plot data collected in 2017. The temporal time difference between the ‘LiDAR AGB model’ and the Sentinel-2 scenes may introduce error into the Random Forest regression; however, the areas of the reserve covered were established to be free from logging and fire damage in this time period, with no obvious structural disturbance.
The Random Forest algorithm is widely used in the remote sensing community [
37,
38,
64,
65] due to its capacity to handle potentially noisy and highly correlated variables often associated with remotely sensed datasets [
63,
66]. Previous research [
32,
40] found the Random Forest method to produce superior results in comparison with other approaches, such as Support Vector Machine, k-nearest neighbour, and Gaussian processes.
Random Forest is an ensemble learning method based on classification and regression trees, this research follows the latter using the ‘Ranger’ package [
67] within R environment software (version 3.5.2). Random Forest regression operates by constructing a large number of trees, i.e., a forest, by randomly selecting the predictor variables from the input dataset and then calculating a mean prediction from all individual trees grown. One-third of the input data, termed the out-of-bag (OOB) sample, are left out of the bootstrap sample and are used to calculate an unbiased estimate for the retrieval error (OOB error). Within the Random Forest regression, three parameters need to be optimised: (1) ntree, the number of trees grown; (2) mtry, the number of different predictors sampled at each node; and (3) nodesize, the minimal size of the terminal nodes of the trees.
In addition to prediction, Random Forest estimates variable importance measures to rank variables by their predictive importance. The ‘Ranger’ package computes the permutation importance of a variable, calculated as the difference in prediction performance before and after permuting the values of the variable averaged over all trees built by the Random Forest. In each tree, OOB data are included in the variable importance calculation. Importance values are ranked and used to determine the predictor strength and influence on the Random Forest model performance. This information can then be used to tailor predictor selection by eliminating low-ranking predictor variables to produce the best predictive power and help interpret the final model.
An independent validation was performed with the test data (30% of the ‘LiDAR AGB model’) unseen by the model. The following statistical parameters were considered for the independent validation: R square (R2), adjusted R square (Adj-R2), Root Mean Square Error (RMSE), Normalised Root Mean Square Error (NRMSE) that results in a percentage (%RMSE), and Bias.
2.7. Wall-to-Wall AGB for NSPSF
Wall-to-wall peat swamp forest AGB maps over the entire reserve were generated using the Random Forest model for both the 2015 and 2018 scenes. The use of the Random Forest model, built from the subset Sentinel-2 2015 variables and ‘LiDAR AGB model’, to estimate AGB across the entire site for the 2018 scene was deemed appropriate as the training areas covered a good representation of the study site but importantly also covered a large area of all vegetation types. Thematic AGB maps were created at a 10 m spatial resolution (0.01 ha).
2.8. 2018 AGB Validation
The predictive ability and transferability of the Random Forest model for estimating AGB across the entire reserve for 2018 was assessed using an independent validation dataset of field inventory plot data collected in 2018. A one-to-one relationship between measured and predicted AGB values was fitted. Common statistical parameters were computed—R square (R2), adjusted R square (Adj-R2), Root Mean Square Error (RMSE), Normalised Root Mean Square Error (NRMSE) that results in a percentage (%RMSE), and Bias.
2.9. AGB Change Analysis
The NSPSF AGB maps generated for 2015 and 2018 (10 × 10 m pixel resolution) were used to calculate an AGB change map by subtracting the 2018 AGB map from the 2015 AGB map. The resulting AGB change map was visually analysed and interpreted with reference data acquired from the Selangor State Forestry Department and local community and conservation groups.
The following information was provided:
Dates and locations of drain-blocking activities;
Dates and locations of replanting programmes;
Dates and locations of large-scale fire events;
Locations of human hydrological interventions, e.g., clay bunds/dyke installations;
Recent changes in land adjacent to NSPSF, e.g., road upgrades.
Interpretation of NSPSF AGB change between 2015 and 2018 was also supported by the IMP documentation and personal knowledge of the reserve obtained from frequent field campaigns in NSPSF.
4. Discussion
Planning and management to restore and conserve tropical peat ecosystems require spatially explicit and timely estimates of AGB, in particular for tracking the impacts of any restoration interventions [
6]. As one of the largest remaining tropical peat swamp forests in Peninsular Malaysia, NSPSF has been recognised for its global significance as a carbon store [
68,
69,
70]. As a result, its ecosystem health and security are of immense importance. The predictive Random Forest model and AGB maps presented in this research provide a quantitative tool for peat swamp forest management in NSPSF reserve, with the potential to help support current and future multi-stakeholder community-based restoration activities [
71]. Additionally, it affords the ability to isolate key areas of degradation (both anthropogenic and natural), which can aid in prioritising and formulating sustainable management strategies.
The time period under investigation in this research (2015–2018) marks a focused period of restoration and regeneration efforts in the NSPSF reserve. The results from the AGB change analysis demonstrate an overall increase in tropical peat swamp AGB during these years, where the total amount of estimated AGB stored in NSPSF increased from 19.3 Tg in 2015 to an estimated 19.8 Tg in 2018, totals comparable to other AGB estimates conducted in tropical forests of similar size and density in Southeast Asia [
8]. Despite this positive result, areas of concern remain within the reserve where continued AGB losses occurred, in particular toward the southeast corner of the reserve. The major factors implicating tropical peat swamp forest recovery and regeneration for NSPSF are drainage (both within the reserve and as a result of land use adjacent to the reserve) and large-scale fire events. Drainage from the estimated 500 km network of canals within the reserve, as well as a result of adjacent oil palm plantation development and ex-mining activities, continues to stress this vulnerable ecosystem [
44]. Altering the natural balance of the water table, in combination with frequent large-scale fire events [
72], can drive a negative feedback loop in tropical peatland ecosystems, affecting their function and resilience to further degradation [
1,
73]. The current canal-blocking activities and water table management efforts [
74,
75] must be monitored and extended throughout the entire reserve. Only by rewetting the peat can hydrological functions be restored, peat oxidation decreased, and fire risk reduced in NSPSF. In light of our results, it is encouraging to learn that the Selangor State Government, with technical assistance from the Global Environment Center, is currently updating and extending the IMP to 2033. The IMP revision will draw on the results from this research (Faizal Parish, pers comm May 2024). It is important to note that recovery can be a slow process, particularly when analysing AGB trends and dynamics at this scale, as large-scale fire events will have a short but dramatic impact, which can form a dominant marker in time series analysis when compared with long-term forest regeneration. It is therefore important to have more long-term (+5 years) studies on tropical peat swamp restoration which correspond with typically longer-term IMPs (i.e., 10 years) to analyse which methods are the most successful in tropical peat recovery [
73,
76]. This is an area of future work in which remote sensing technology could be usefully integrated.
The Random-Forest-based method used in this research shows great potential to upscale AGB estimates, derived from the ‘LiDAR AGB model’, to larger landscape-scale areas with Sentinel-2 spectral and textural variables. The variable importance rankings were used to assess the influence of Sentinel-2 spectral and textural inputs on the predictive performance of the Random Forest model. This research found a tailored variable selection approach improved predictions of AGB, this is in line with previous research applying the Random Forest algorithm to remotely sensed data [
77,
78]. The inclusion of textural variables had the greatest importance for modelling AGB in tropical peat swamp forests using optical satellite data (
Figure 3). Texture values in an image allow the spatial heterogeneity to be evaluated, this can provide information on the distinct structure and development stages of forest stands. In particular, textural variables applied to Sentinel-2 bands positioned in known green vegetation spectral features, for example, the Red-Edge (bands 5, 6, and 7) and narrow near-infrared bands (band 8a), improved AGB estimation. The advantage of Sentinel-2 bands covering the Red-Edge zone has been noted in a number of forestry studies [
22,
52,
79]. Additionally, vegetation indices incorporating Red-Edge bands proved to be important for modelling AGB. Laurin et al. [
80] suggested that the use of vegetation indices incorporating the Red-Edge bands may help mitigate the saturation issues associated with the use of traditional vegetation indices in the tropics. A potential opportunity to improve AGB modelling in future work could be the inclusion of textural variables applied to vegetation indices incorporating Red-Edge bands, combining the benefits of both approaches. Past research has successfully explored the use of combining both spectral and spatial techniques, namely texture measurement and image ratio [
81].
It is well documented that the spatial distribution and sample size of training data have considerable influence on the accuracy of model predictions [
32,
82,
83]. Here, the research adopted the results from a model developed by Brown et al. [
9] as training and validation data in replacement of traditional field inventory plots. Errors introduced in the model development by Brown et al. [
9] will inevitably influence the overall performance of the Random Forest model and AGB map outputs produced in this research. Although it must be stressed that the ‘LiDAR AGB model’ sample size and heterogeneity are considerably larger and better distributed than could be attained from a field-based programme covering the same spatial and temporal scope, it is also important to note that both the ‘LiDAR AGB model’ dataset and the forest inventory plot estimated AGB values are reference datasets and do not represent actual measurement of AGB; only destructive sampling techniques can afford this. As previously discussed, potential errors may also have been introduced into the AGB models through temporal discrepancies between the ‘AGB LiDAR model’ and the Sentinel-2 2015 scene, and additionally by geo-correlation issues between datasets. Combined, significant deviations from ‘true’ AGB values may occur.
Remote sensing has played a vital role in practical conservation management [
84,
85,
86], establishing baselines for the extent and characteristics of ecosystems and their services, as well as quantifying degradation and subsequent recovery associated with both natural and anthropogenic events or processes. There are a number of trade-offs associated with the use of different remote sensing technologies and sources (such as data acquisition cost, coverage, and platform considerations), along with their associated level of data complexity and processing expertise. Although alternative remote sensing technology such as LiDAR has been proven to be superior in AGB estimation in previous research [
32], optical spaceborne data remains a practical option for conservation management research applications [
51,
87].
Spaceborne optical data such as the Landsat series have long served as a major data source for ecosystem assessment and management [
88,
89]. The remote sensing community is in a golden age of spaceborne sensor technology, in particular for ecosystem structural mapping and biomass prediction [
90], with the next generation of spaceborne optical sensors (e.g., ESA’s Sentinel-2 and NASA/USGS’s Landsat-9), LiDAR profilers (NASA’s ICESat-2, NASA’s GEDI, and JAXA’s MOLI) and SAR sensors (e.g., ESA’s BIOMASS, NASA/ISRO’s NISAR, and ESA’s Sentinel-1) set to provide a diverse range of datasets specifically designed for AGB assessment at a level of detail and temporal frequency never recorded before. Additionally, opportunities in high/very high spatial resolution satellite data have emerged from constellations of microsats [
91,
92]. These novel sensors operate at a low orbit and a comparatively low cost and thus can be deployed in large numbers. The performance of microsats is rapidly evolving [
93,
94], for example, Planet Labs has deployed a constellation of 100+ Doves microsats [
95]. With a daily 3–5 m data resolution, the Planet Labs constellation has been flagged as an important data resource for assessing AGB, carbon stocks, and emissions at a high spatial and temporal frequency [
96]. The field of remote sensing continues to evolve and revolutionise the way we view Earth’s surface, leading to countless advances at the ground level in both scientific knowledge and practical application. To this end, supporting data are required to tackle global issues such as sustainable land recovery and conservation (SDG 15).