Next Article in Journal
GIDNet: Infrared Small Target Detection Network Based on Gradient-Intensity Decoupled
Previous Article in Journal
A Thresholded NDVI-AUC Metric from Multi-Source Optical Time Series for Mapping Surface Soil Salt Content in Vegetated Coastal Areas
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Remote Sensing
Volume 18
Issue 10
10.3390/rs18101523
remotesensing-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Characterizing Savanna Tree Canopy Heights Using GEDI and Spatially Continuous Multi-Source Data at a Landscape Level

by
Xiao Ma
1,
Yajie Qu
2,3,
Meiyuan Chen
2,3,
Guang Zheng
2,*,
Chi Xu
1 and
Xiaoxuan Li
4
1
College of Life Sciences, Nanjing University, Nanjing 210023, China
2
International Institute for Earth System Science, Nanjing University, Nanjing 210023, China
3
Jiangsu Provincial Key Laboratory of Geographic Information Science, Nanjing 210023, China
4
Department of Geography and Geoinformation Science, George Mason University, Fairfax, VA 22030, USA
*
Author to whom correspondence should be addressed.
Remote Sens. 2026, 18(10), 1523; https://doi.org/10.3390/rs18101523
Submission received: 8 April 2026 / Revised: 7 May 2026 / Accepted: 8 May 2026 / Published: 12 May 2026
(This article belongs to the Special Issue Satellite LiDAR for Vegetation Structure and Biomass Estimation: Advances with ICESat-2 and GEDI)
Browse Figures
Versions Notes

Highlights

What are the main findings?
  • Develop an optimized workflow for the filtering of high-quality GEDI footprints.
  • Establish a mapping framework for accurately estimating savanna tree canopy height.
What are the implications of the main findings?
  • The proposed method enables the filtering of accurate footprints from GEDI L2A products without relying on high-precision reference data.
  • Savanna tree canopy height prediction requires modeling distinct from that for forests.

Abstract

Accurately mapping tree canopy heights of savanna ecosystems, which account for around 20% of the terrestrial land surface, is of great importance for global biomass estimation, carbon cycling, and biodiversity. The spaceborne lidar of Global Ecosystem Dynamics Investigation (GEDI) has great potential for measuring tree canopy heights in sparse savanna ecosystems due to its implicit three-dimensional structural information. However, the accuracy of the GEDI system may be affected by the random geolocation errors. In this study, we aim to develop a reliable method to mitigate the impact of low-quality and position-biased GEDI footprints. Then we generated 30-m resolution wall-to-wall mapping of tree canopy heights for 2020 by combining GEDI L2A footprints with spatially continuous multi-source information in the Kruger National Park, South Africa. Moreover, we explored the explanatory ability of multi-dimensional features derived from optical, radar, topographic, and artificial intelligence-based images and conducted a comparative analysis of relevant products. Validation results confirmed that integrating quality indicators, incorrect ground elevation estimation assessment, and optical and radar features could significantly improve the accuracy of GEDI-based tree canopy height estimation in savannas (i.e., Pearson’s r = 0.51, RMSE = 3.88 m, N = 6276). Compared to existing products, the model trained on comprehensively filtered footprints exhibited higher agreement with reference canopy height model data and lower estimation errors (i.e., Pearson’s r = 0.66, RMSE = 4.09 m, N = 10,469). We also found that features incorporating red-edge bands exhibited higher explanatory ability. This study showcases GEDI-based mapping of savanna tree canopy heights and provides a foundation for future large-scale research on savanna ecosystems.

1. Introduction

Savanna ecosystems cover 20% of the Earth’s land surface and account for one-third of terrestrial net primary production globally [1]. They are three times larger in area than closed forests in Africa [2,3], and typically coexist with grasses, shrubs, and trees in tropical and subtropical regions [4,5,6]. The obvious heterogeneous spatial distribution patterns of tree cover, fire regimes, and water or nutrient availability result in diverse ecological niches for mammalian species [7], providing indispensable habitats for the majority of the world’s livestock and wildlife [8]. Furthermore, savanna ecosystems serve as the economic foundation for local communities by providing opportunities for livestock grazing, agricultural land, and sources of timber and non-timber products [9,10].
Wildfires, domestic and wild animals, climate and soil conditions continuously affect and shape the landscape pattern of savanna tree-grass coexistence [11,12]. These dynamic changes further alter biodiversity, energy budget, and carbon and water cycles in the local ecosystem. Due to the extensive coverage of savannas on the global terrestrial land surface, they have a significant impact on the Earth–atmosphere feedback [13,14]. The spatial configurations and three-dimensional (3D) structures of savanna ecosystems, especially tree canopy height, are essential and key parameters to better estimate aboveground biomass and wood volume [15,16], monitoring the dynamics of tree-grass interactions [17], investigating fire frequency [18], and studying animal behavior and reproduction in savannas ecosystems in terms of ecological and socio-economic values [19,20].
The prevalent approach for measuring savanna canopy height is the field-based method, which is labor-intensive and time-consuming, thereby limiting its scalability. The development of photogrammetric, terrestrial, and aerial laser scanning data has enabled successful characterization of savanna tree canopy heights [21,22,23]. However, their high costs make them challenging to generalize on a large scale [24].
The emergence of spaceborne light detection and ranging (lidar) enables the direct acquisition of 3D structural information of vegetation canopies through discrete global sampling. However, early spaceborne lidar missions faced substantial limitations in resolving savanna canopy structure. NASA’s ICESat GLAS, launched in 2003, has a limited ability to retrieve savanna canopy heights due to its relatively large footprint diameter of 70 m. In 2018, ICESat-2 was launched with the original design of measuring ice sheet and sea ice thickness with a green light wavelength (i.e., 532 nm). However, the low reflectance of vegetation in this wavelength limits its ability to extract savanna canopy height [25]. In contrast, deployed on the International Space Station in 2018, the Global Ecosystem Dynamics Investigation (GEDI) is specifically designed for vegetation monitoring. It samples the ground surface with a finer 25 m diameter footprint between 51.6°N and 51.6°S, capturing full-waveform returned intensities data to characterize the vertical profile of forest canopies [26].
Recently, Li, et al. [27] demonstrated that GEDI could effectively capture variations in savanna canopy heights based on the full waveform information within each footprint. Given that the observations of savanna canopy heights at the footprint level are discrete, it is necessary to develop an extrapolation method to create a wall-to-wall map by combining GEDI-based footprint level results with various spatially continuous remotely sensed data. Optical data and synthetic aperture radar (SAR) are the primary tools used to scale up lidar-derived canopy heights [28,29], with machine learning as the predominant approach for extending these measurements across large areas [30,31]. Traditional models like Random Forest (RF) and Support Vector Machine (SVM) rely on manual feature extraction, such as multispectral indices and texture features, to process optical data effectively [32]. Incorporating time-series phenological indices has further refined canopy height estimates by capturing seasonal changes in forest dynamics [33]. However, optical sensors have inherent limitations: they cannot directly capture vertical forest structure and are susceptible to saturation and cloud cover interference [34]. SAR complements optical data by providing structural information independent of atmospheric conditions. Recent studies combining SAR data from Sentinel-1 and PALSAR-2 with optical imagery have shown improved accuracy in canopy height mapping [35,36]. Moreover, AlphaEarth Foundations provided a globally multi-dimensional dataset with a 10-m resolution by synthesizing petabytes of Earth observation data based on the new artificial intelligence (AI) model [37]. Integrating these datasets may enhance model accuracy by providing complementary information.
Current studies on spatially continuous canopy height mapping based on GEDI data face challenges caused by geolocation uncertainty of the GEDI system. Tang, et al. [38] confirmed that most of the geolocation errors of GEDI V2 were less than 10 m with a study in the United States, which significantly improves the accuracy of the canopy height extracted from GEDI data. However, in savanna ecosystems characterized by high heterogeneity in vegetation distributions, the current version of GEDI data may still introduce modeling errors arising from horizontal positioning uncertainty. Previous researchers mostly followed simple rules to filter GEDI footprints [39]. Schleich, et al. [40] computed the error map between GEDI ground estimation and DEM for each footprint and employed a flow accumulation algorithm to identify the optimal positions. Moudrý, et al. [41] proposed an effective approach for finding accurate GEDI footprints by combining the quality flags and digital elevation model (DEM) difference from TanDEM-X. Investigating systematic geolocation errors of GEDI footprints remains a challenge due to the high cost and difficulty in obtaining essential precise point cloud data or high-resolution DEM data [42,43]. Therefore, in the savanna ecosystem with a discrete canopy distribution, it is critical to develop a comprehensive filtering method for mitigating the impact of horizontal geolocation errors of the GEDI system.
In this study, we address a key limitation in GEDI-based canopy height mapping by developing a comprehensive GEDI footprint filtering framework tailored to heterogeneous savanna ecosystems. Our approach integrates multiple indicators, including GEDI quality metrics, terrain consistency, and multi-source remote sensing features based on local environmental conditions. The specific objectives are to: (1) develop a reliable way to select accurate GEDI footprints in savanna; (2) map spatially continuous savanna tree canopy heights from a footprint to a landscape level; and (3) evaluate and analyze our proposed methods in filtering data and mapping results.

2. Study Area

Kruger National Park (KNP), established in 1926, is located in the northeast of South Africa. Covering approximately 20,000 km2, the park has a long, narrow shape running from north to south (Figure 1). The terrain of KNP is generally flat, with elevations ranging from 200 to 840 m above sea level, averaging around 260 m. The mountainous areas are mostly located along the eastern boundary, formed by the Lebombo Mountains [44].
The park has a hot, semi-arid climate with distinct rainy and dry seasons. The rainy season typically lasts from September to May of the following year [45]. There is significant variation in rainfall across the park’s north-south span, with the southern region receiving the highest amount of rainfall and the central area experiencing the least. The average annual rainfall ranges from 350 mm to 750 mm from south to north [46]. In January, the average temperature is around 30 °C, while in July, it drops to around 23 °C. The park can reach extremely high temperatures of up to 47 °C [47].
The park is classified as a deciduous savanna with varying vegetation density [48]. The Olifants River divides the park into north and south areas. Remote sensing revealed significant variability in tree cover throughout the park (Figure 1a). Higher tree cover is concentrated in the southern and southwestern regions, where precipitation is relatively higher, while tree cover becomes sparser moving northward. The northwestern area has sparse vegetation, while the northeastern area is dominated by multi-stemmed Colophospermum mopane shrubs that reach a height of 1–2 m. The southwestern part of the park is well-wooded and receives higher rainfall, while the southeastern part is primarily grassland and heavily grazed [49].

3. Materials and Methods

3.1. Data

3.1.1. GEDI

The GEDI instrument, which is installed aboard the International Space Station, uses three lasers (two full power lasers and one coverage laser) to conduct detailed 3D measurements of the Earth’s surface [26]. The Version 2 (V2) data, which was released in 2021, offers improved geolocation and better selections of algorithm setting groups compared to the Version 1 data [50]. We obtained the GEDI V2 L2A product covering KNP from the Google Earth Engine (GEE) platform from October 2019 to April 2020. The L2A product includes ground elevation, and relative height (RH) metrics, which are calculated directly from the waveform return. RH metrics represent the elevation at which a specific quantile of returned energy is achieved [51]. We used the RH98 metrics to represent the tree canopy height because it contains fewer outliers and is less sensitive to noise compared to RH100 [52]. The GEDI laser pulse width of 15.6 ns makes it unreliable to estimate canopy heights of shrubs below 2.35 m. Therefore, we excluded all footprints where RH98 is under 2.35 m [27].

3.1.2. Optical Data

Sentinel-2 is a multispectral operational imaging mission that provides high-resolution, multi-spectral imaging with wide swath coverage. Its frequent and systematic monitoring significantly contributes to land monitoring services (e.g., vegetation dynamics and land surface classification) [53,54]. The atmospherically corrected L2A product in 2020 was retrieved from GEE, and clouds and shadows were masked using the Sentinel-2 quality bands. All 20 m bands were resampled to 10 m resolution with bicubic interpolation.
Vegetation indices such as Enhanced Vegetation Index (EVI), Normalized Difference Vegetation Index (NDVI) [55], Modified Soil-adjusted Vegetation Index (MSAVI) [56], Blue Chromatic Coordinate (BCC) [57], Excess Green Index (ExG) [58], Green Chromatic Coordinate (GCC) [57], and Green Leaf Index (GLI) have been widely used for predicting canopy height [39]. We calculated these indices on Sentinel-2 data to analyze vegetation patterns. The Sentinel-2 satellite’s three red edge bands were designed for vegetation monitoring, and band B8A is sensitive to total chlorophyll and biomass [59]. For these four bands, we calculated the Canopy Chlorophyll Content Index (CCCI) [60], MERIS Terrestrial Chlorophyll Index (MTCI) [61], Sentinel-2 Red-Edge Position (S2REP) [62], Transformed Chlorophyll Absorption in Reflectance Index (TCARI) [63], and NDVIs with these four bands replacing NIR and Red bands. Table 1 lists all vegetation indices derived from Sentinel-2 and their formulations.

3.1.3. Radar Data

The Sentinel-1 mission provides data with a 10-m resolution from a dual-polarization C-band SAR instrument. The short C-band wavelength is highly sensitive to grasses and low vegetation [64]. Sentinel-1 GRD collection on GEE was thermal noise removed, radiometrically calibrated, and terrain-corrected using Sentinel-1 Toolbox. For this study, we used the VH, VV and the ratio of them in 2020. The VHVV index has been commonly utilized to identify forested areas and individual trees [65].
The 25 m L-band PALSAR-2 ScanSAR is normalized backscatter data of PALSAR-2 broad area observation mode with an observation width of 350 km. The low return of grassy vegetation at both HH and HV polarization allows for better recognition of trees in savannas [64]. The collection on GEE was ortho-rectified and slope-corrected. Indices of HH, HV, and HHHV in 2020 were collected and sampled to a 30 m grid size based on the nearest neighbor method.

3.1.4. AlphaEarth Data

AlphaEarth Foundations (AEF) is an embedding-based model that constructs a unified geospatial representation by integrating spatial, temporal, and multi-source measurement contexts [37]. This framework enhances the development of precise and efficient mapping and monitoring systems across local to global scales. The Satellite Embedding Dataset, generated by the AEF model, provides a globally spatially consistent, analysis-ready geospatial embeddings spanning from 2017 to 2024. The dataset provides global coverage of land surfaces and shallow water bodies at a 10-m spatial resolution, with each image consisting of 64 bands (Table 1). The AEF data for 2020 was collected through GEE.

3.1.5. Topographic Information

The Shuttle Radar Topography Mission (SRTM) aimed to acquire digital elevation model (DEM) on a near-global scale [66]. Our study used the SRTM V3 product provided by NASA JPL at a resolution of approximately 30 m. We extracted slope and aspect from elevation data using GEE. Topography plays a significant role in shaping local nutrient and water conditions, which in turn affects tropical forest structures and compositions [67].

3.1.6. Mask Data

To filter GEDI footprints and validate our canopy height map, we utilized a 30 m resolution Africa tree cover map (Figure 1b) [68]. This dataset provides a detailed delineation of tree crown cover across continental Africa, derived from 3 m PlanetScope imagery in 2019 using a deep-learning segmentation framework based on an extended U-Net architecture [69]. The resulting product represents the percentage of tree cover aggregated to a 30-m spatial resolution. This fine-scale mapping allows for consistent detection of tree canopies across both forested and non-forested landscapes, including savannas and woodlands. We masked out all areas without tree cover (i.e., pixels with tree cover = 0) to focus our analysis on the tree-canopy layer, which ensures that GEDI footprints are evaluated only where woody vegetation is present.

3.1.7. Reference Data

Previous studies have shown that a high-resolution digital surface model (DSM) can serve as an alternative to lidar for estimating canopy height and forest biomass [70]. We derived the reference CHM data by subtracting the DEM from the DSM. The DSM and DEM were obtained from color-infrared aerial imagery using stereo-matching algorithms [23]. The Leica Digital Mapping Camera (DMCIII) was used to collect aerial images in September/October 2018, employing a CIR channel combination of the NIR, red and green bands at an altitude of approximately 5500–6000 m above ground. The digital terrain model (DTM) was validated using Global Navigation Satellite System (GNSS)-based ground survey points, while very high-resolution UAV-lidar data was used for DSM validation. The horizontal accuracy of the ortho-mosaics was confirmed using triangulation stations, and the final products demonstrated excellent agreement with R2 values of 0.99 [23]. Given that the original CHMs derived from 1 m and 5 m resolution data still displayed significant topographic variations, we combined the high-quality segments from both resolutions via a mosaicking process to produce a consolidated CHM for subsequent accuracy evaluation. The aggregation of the 1 m and 5 m resolution CHMs to a unified 30 m resolution was performed using the 98th percentile and maximum value methods, respectively. As shown in Figure 1b, most trees are relatively short, with only a small proportion exceeding 10 m in height. The masked CHM data for the KNP region (i.e., tree cover > 0 and height > 2 m) yielded a mean value of 5.75 m and a 99th percentile value of 18.47 m. Based on the multi temporal canopy height products in KNP from Filippelli, et al. [71], we quantified the canopy height changes between 2018 and 2020 across the study area. The statistical results, with the 25th and 75th percentiles recorded at −0.30 m and 0.23 m respectively, indicate that no major disturbances occurred during this period, thereby meeting the requirements for accuracy assessment.

3.2. Features Calculation

To fully utilize the phonological characteristics from the time-series data and exploit the spatial details provided by relatively finer-resolution images with a 10-m grid size, we constructed explanatory features by calculating temporal statistics and aggregated measures. Multitemporal statistical features, calculated from optical and SAR time series datasets, contain rich spatiotemporal information and have been proven to contribute to modeling canopy heights [24]. The available observations of Sentinel-2, Sentinel-1, and PALSAR-2 in 2020, are shown in Figure 1d–f. These data presented a sufficient number of available observations to support the computation of statistical features for each pixel based on time series data. We calculated the mean, variance, percentage quantiles (P10, P25, P75, and P90), and interval subtracts of P100-P75, P75-P25, and P25-P0 for raw bands, all spectral indices, and all radar indices in this study. All features with a 10-m resolution (i.e., Sentinel-2, Sentinel-1, and AlphaEarth) were sampled to a 30-m grid size with mean and standard deviation values (Table 1). Additionally, topographic information was incorporated to enrich the explanatory features. Finally, there were 716 features derived from multi-source data sources (Table 1).

3.3. Comprehensive Filtering Framework for GEDI Data

Most studies employ a simplified filtering process based on GEDI quality flags to identify high-quality footprints. However, in open-canopy savanna environments, geolocation errors of GEDI footprints can lead to issues like mismatches between relative height and remote sensing indices. To minimize their influence, we developed a comprehensive filtering framework by combining GEDI attributes, elevation bias, and optical and radar features (Figure 2). First, we used the simplified filtering process recommended by the GEDI level 02 user guide [50] to obtain high-quality data and to ensure that the footprint was on vegetation. (1) Set quality_flag to 1, degrade_flag to 0, and sensitivity greater than 0.9; (2) ensure that tree cover is greater than 0.
Subsequently, we filtered the footprints based on the elevation bias (BiasDEM) calculated by leveraging the Tandem-X DEM and the Elevation_lowestmode metric provided in the GEDI L2A product. For all the footprints with RH98 ranging from 0 to 30 m, we conducted a simple linear regression between the median elevation bias and median interval value of tree canopy height computed within each 5 m interval. This range could cover the majority of tree canopy heights observed across the KNP area. The 1.5 interquartile range (IQR) method was then applied to the residuals derived from the fitted regression to retain footprints within the threshold range. This approach effectively reduced the impact of footprints with substantial elevation estimation errors [72].
We finally utilized optical and radar features to filter out potentially position-biased footprints. For optical features, we identified a strong positive correlation between the NDVI_B87_P75_std index and tree canopy height by comparing the relationships between various vegetation indices and RH98 metrics. Following a similar procedure, we also fitted a regression line for the highly correlated optical index and calculated the estimation residuals. These residuals were then used in an IQR-based filtering approach to eliminate biased footprints. Footprints with RH98 < 10 m were kept if their residuals fell within the upper threshold, while those with RH98 ≥ 10 m were excluded if their residuals fell below the lower threshold. Considering the complex relationship between remote sensing features and tree canopy height, we adopted a conservative approach to avoid over-filtering, removing only the most apparent outliers. Additionally, we also observed that the NDVI_B87_P75_std index yields high values in areas characterized by high tree cover despite low canopy height. Therefore, we preserved footprints with RH98 < 10 m and tree cover ≥ 50% even when their residuals exceeded the upper threshold.
For radar features, we employed the VV_Variance_mean index for the further filtering process (Figure 2). This distribution of VV_Variance_mean showed clustering in the medium-to-high range and dispersion in the low range (i.e., tree canopy height < 10 m). Therefore, we applied this characteristic to identify additional potentially position-biased footprints with height mismatches. Using a segmented approach, we applied the 1.5 IQR method to filter outliers, specifically preserving footprints with RH98 < 10 m and tree cover < 25%. The spatial distribution of footprints after this comprehensive filtering process is shown in Figure 1c.

3.4. Random Forest-Based Tree Canopy Height Estimation

A Random Forest regression model was utilized to predict the tree canopy height in KNP. The model was implemented using the ranger package in the R language [73]. The Random Forest is a non-parametric statistical estimator that fits a number of “weak” regression trees on sub-samples of the dataset, and employs averaging to enhance prediction precision and manage overfitting [74]. The Random Forest algorithm is a well-established method for integrating lidar data with other remotely sensed data for large-scale canopy height mapping [24,35,75,76]. This method was chosen because it can uncover complicated non-linear relationships, allows for the calculation of feature importance, and can efficiently run on large datasets [77].
To validate the effectiveness of the proposed filtering framework, we built regression models using sifted footprints after simply filtered by GEDI attributes and by our proposed comprehensive filtering (Figure 2). Two parameters, mtry and ntree, were important in the modeling process and were set to one-third of the candidate features and 1000, respectively. Since most of the RH98 values from the filtered footprints were less than 10 m, severe skewness in the data distribution could affect the accuracy of estimating the higher values during training models. To mitigate this issue, we first performed stratified sampling on footprints within the 0–5 m and 5–10 m height intervals. Subsequently, weighting coefficients were calculated at 5 m intervals across the 0–30 m range based on the RH98 values of the filtered footprints, using the inverse frequency weighting method. The calculation formulation is given below. Ni represents the number of footprints falling within the i-th interval.
W i   = ∑ i 6 N i N i
To enhance training efficiency and prevent overfitting, we selected explanatory features by calculating mean values of feature importance measurements across multiple training iterations. Using the variable importance measure provided by the ranger package, we calculated the mean importance value and retained features exceeding a threshold of 0.01. Finally, all filtered footprints were used to train the Random Forest model.

3.5. Accuracy Assessment

We conducted accuracy assessments using the reference CHM data on the comprehensively filtered GEDI footprints, the Random Forest estimation model, and the mapping results. The evaluation employed two commonly used indicators: the Pearson’s correlation coefficient (r) and root mean square error (RMSE). A higher r value indicates more agreement between the estimated values and the reference data, while a lower RMSE reflects a smaller estimation error. The performance of the trained tree canopy height estimation models was evaluated through 50 random train-test splits (i.e., 75% for training and 25% for testing). The mean values of the two indicators were calculated. We further evaluated the performance of the trained tree canopy height estimation models using the GEDI RH98 relative height metric.
Subsequently, we conducted an inter-comparison between our mapping results and four existing products within the KNP region. Filippelli, et al. [71] provided a time series dataset of canopy heights at a 30-m grid size for the Greater Kruger National Park region of South Africa, spanning the period from 2007 to 2022. This dataset was generated through the calibration of models integrating GEDI RH98 metrics with multi-source data, including Landsat data, PALSAR data, topographic information, and soil data. Their 2020 dataset was included in our comparative analysis. Potapov, et al. [24] produced a global forest canopy height map in 2019 by combining GEDI footprints with Landsat analysis-ready data. Lang, et al. [78] created a 10-m resolution canopy height map for 2020 using a deep learning model trained on GEDI footprints and Sentinel-2 images. Additionally, a high-resolution global canopy height product with a 1-m grid size provided detailed structural insights into global trees around 2020 [79]. To ensure comparability of spatial resolution, the 10-m and 1-m global products were resampled via maximum value aggregation to match the spatial resolution of our map.

4. Results

4.1. Comprehensive Filtering Framework Assessment

Figure 3 illustrates the consistency between the RH98 metrics derived from two different filtering methods and the reference CHM data. After applying the comprehensive filtering framework, the correlation relationship between the RH98 metrics and the reference data showed significant improvements, increasing from 0.43 to 0.51. Furthermore, the RMSE value decreased from 4.10 m to 3.88 m. Compared to Figure 3a, Figure 3b demonstrates that the comprehensive filtering framework may effectively remove GEDI footprints exhibiting significant positional errors. Several discrete footprints that were significantly overestimated or underestimated were effectively filtered out, without markedly altering the data distribution. The distribution of most footprints was concentrated on both sides of the reference line.

4.2. RF Models in the KNP

The statistical results indicated a certain increase in the proportion of footprints with medium-to-high values after sampling. The overall data distribution trend remained largely unchanged, with minimal impact on statistical metrics. The mean values of RH98 metrics of the simply filtered footprints and footprints filtered by the comprehensive filtering framework increased from the original 5.69 m and 5.57 m to 6.22 m and 5.95 m, respectively (Figure 4a,b). To mitigate the influence of training data volume on both models, the final number of footprints used for modeling was kept relatively consistent.
Across 50 splits of training and test sets, we evaluated the accuracy of the Random Forest estimation model based on the RH98 metrics and reference CHM data. The results demonstrated that the model trained on comprehensively filtered samples achieved superior accuracy compared to that trained on simply filtered samples. For the comprehensively filtered samples, the mean values of Pearson’s r based on the RH98 metrics and reference CHM data were 0.72 and 0.69, respectively, whereas for the simply filtered samples, the corresponding values were 0.70 and 0.67 (Table 2). Figure 4e,f display scatter plots of the predicted values from the estimation models against the validation CHM data. The model trained on comprehensively filtered samples demonstrated stronger agreement with the reference CHM data and was accordingly selected as the final mapping model, achieving a mean Pearson’s r of 0.69 and RMSE of 3.78 m. Figure 4g illustrates that the model trained on comprehensively filtered samples achieves lower RMSE values across multiple tree cover intervals compared to the model trained on simply filtered samples, particularly within the low canopy cover ranges (i.e., tree cover < 50%).
Figure 5 illustrates the mean variable importance of the features retained for constructing the Random Forest model after screening, with a total of 140 features preserved. Red indicates high explanatory power, while blue indicates low explanatory power. Optical features served as the primary explanatory variables, among which NDVI-related indices calculated using the red-edge bands demonstrated superior explanatory power (Figure 5b). The other features exhibited moderate explanatory capacity, though only HV_mean, VV_P75_mean, and VV_Mean_mean showed relatively high importance (Figure 5e,f). A relatively larger number of explanatory features were retained from the AlphaEarth dataset, though few of them displayed high explanatory power (Figure 5a). Due to the insufficient explanatory power, topographic features were not retained in the final model.

4.3. KNP Canopy Height Maps

Figure 6 presents the wall-to-wall canopy height map at a 30-m resolution for KNP in 2020, which shows relatively dispersed vegetation with a lower height of 3–7 m. Medium to tall trees were concentrated along the rivers and their tributaries, while the northern boundary area adjacent to the Limpopo River accommodated a significant amount of vegetation over 15 m in height.
We randomly selected four sample plots from north to south to further show the accuracy of the mapping results and their horizontal distribution (Figure 6(b1–b4)). We compared our results (Figure 6b(1I,2I,3I,4I)) with processed CHM reference data (Figure 6b(1II,2II,3II,4II)) and high-resolution imagery from Bing image map (Figure 6b(1II,2II,3II,4II)). The findings demonstrated that our proposed method could effectively present the distribution of tree canopy heights, with variation trends showing strong agreement against reference data while capturing rich spatial details. Figure 6(b1,b4) demonstrate that taller trees are predominantly clustered along riverbanks, whereas the surrounding vegetation is generally shorter and more sparsely distributed. Figure 6(b2,b3), in turn, show an area with discretely distributed trees and a region containing sparse clusters of taller trees, respectively. The comparison of these four enlarged sample points indicated that our results could effectively distinguish between high and low vegetation under different conditions and capture regional heterogeneity in the horizontal direction of the canopy, showing good consistency with reference CHM data and high-precision satellite imagery. Nevertheless, our results demonstrated a systematic overestimation of low-stature trees and potential saturation effects, as evidenced by the constrained distribution of dark blue pixels representing these shorter trees in Figure 6(b2I,b3I).

4.4. Comparison with Existing Products

We quantitatively evaluated our results and existing products using randomly sampled points located outside the GEDI footprint distribution area. The results demonstrated that our product achieved the highest agreement with the reference data and the lowest overall estimation error, with a correlation coefficient of 0.66 and an RMSE of 4.09 m (Table 2). In contrast, maps from Filippelli, et al. [71] product covering the same region and period exhibited significantly weaker performance, with Pearson’s r and RMSE values of only 0.15 and 5.62 m, respectively. Three global forest canopy height products evaluated showed greater uncertainty in sparse grassland areas, with Pearson’s r values ranging from 0.19 to 0.45 and RMSE values between 4.98 m and 6.91 m. We further assessed the accuracy of these maps across different height classes. For tall trees (i.e., height ≥ 20 m), both Filippelli, et al. [71] and Potapov, et al. [24] products exhibited significant underestimation (Figure 7a,c), with RMSE values of 14.59 m and 17.03 m, respectively. Although our results also showed a tendency to underestimate in this height range, the RMSE in the high-height interval was 7.83 m—only slightly higher than that of the Lang product with the RMSE value of 7.25 m. For medium trees (i.e., height < 20 m and height ≥ 10 m), our product achieved the highest estimation accuracy, with a Pearson’s r of 0.49 and an RMSE of 3.17 m. By comparison, the best-performing existing product in this class reached a Pearson’s r of only 0.34 and a minimum RMSE value of 5.76 m. Although our mapping results achieved superior agreement with reference data for shorter trees compared to other products, they exhibited relatively high errors (i.e., RMSE = 4.28 m) due to overestimation. This systematic bias was visually evident in Figure 7f, which showed the error distribution of our mapping results skewed to the right of the reference line. Overall, the estimation distribution of our results was more concentrated, without significant dispersion across the height range (Figure 7e).
To more comprehensively evaluate the trends of estimation residuals from different maps under common influencing factors, we established separate univariate linear regressions between the residuals and tree canopy height, tree cover, and slope. All regression slope coefficients were statistically significant. The results indicated that while the estimated residuals of all products increased with tree canopy height, the increase was less pronounced for our results and Lang, et al. [78] and Tolan, et al. [79] products (Figure 7g). Furthermore, no commensurate increase in residuals was observed with higher tree cover (Figure 7h). In contrast, steeper slopes were associated with elevated residual levels across all mapping results. Notably, the increase in residuals from our results was the most moderate, with a slope coefficient of merely 0.12 (Figure 7i).
Close-up views of four sample plots representing five different tree canopy height maps and the CHM with the same color provided more detailed product comparisons (Figure 8). The maps derived from our method, Lang, et al. [78] and Tolan, et al. [79] products effectively captured the overall trends in tree canopy height variation across the KNP (Figure 8(a1,a3,a5)). In particular, our results better represented sparsely distributed medium or low trees (Figure 8(a1–d1)). In areas where medium and tall trees were concentrated, our approach more realistically reflected horizontal tree canopy heterogeneity and more accurately represented the tree structure of sparse tree-grass landscapes. In contrast, the results from Lang, et al. [78] and Tolan, et al. [79] exhibited a systematic overestimation, which compromised their ability to present fine spatial details in medium-to-tall tree areas (Figure 8(a3,c3,a5,b5)). Additionally, the maps from Filippelli, et al. [71] and Potapov, et al. [24] products performed poorly in savanna areas, underestimating tree canopy heights above 16 m (Figure 8(a2,d2,a4,d4)). Detailed comparisons confirmed that Filippelli, et al. [71] and Potapov, et al. [24] maps were unable to accurately present medium-to-tall tree heights, resulting in substantial underestimation.

5. Discussions

5.1. Effectiveness of Proposed Filtering Method

The assessment results demonstrated that our proposed comprehensive filtering framework could effectively filter out footprints with potential geolocation errors for the GEDI L2A product. While quality control indicators provided a basis for the rapid finding of accurate GEDI footprints, we further developed a comprehensive filtering framework based on the potential impacts caused by position-biased footprints to ensure methodological robustness. Significant discrepancies between the reference DEM data and the lowest ground elevation estimated by the GEDI full-waveform data may indicate inaccuracies in ground detection [39,41,80]. Such errors in ground elevation estimation can be influenced by factors such as terrain slope [80,81], dense vegetation cover [82], and horizontal positional bias [43]. Given that the topographic variation within the study area is generally gentle and vegetation is predominantly sparse, applying strict thresholds based on ground elevation estimation errors may effectively eliminate some footprints affected by horizontal position bias (Figure 2). Even in steep and densely forested landscapes, incorporating a filtering step based on ground elevation error can help remove unreliable GEDI footprints. Horizontal geolocation errors may also lead to mismatches between the relative height and its corresponding spectral or backscatter features. Therefore, exploring highly correlated or characteristic remote sensing indices can help identify anomalously distributed footprints. Establishing suitable filtering thresholds based on these indices further enhances the effectiveness of the filtering processes.

5.2. Feature Selection and Interpretation

In particular, time-series features from Sentinel-2 played a crucial role in canopy height prediction. We believe this is due to the coexistence pattern of trees and grass in savanna ecosystems. Unlike grass, trees exhibit greater growth determinacy, typically turning green before the first rainfall due to deeper root systems and water storage capacity and have a longer growing season [83,84]. Time-series features can capture the phenological differences between trees and grass well. Previous studies employing near-infrared digital repeat photography have successfully tracked the structure and phenology of tree-grass ecosystems to differentiate them [85]. Our results indicated that the variables crucial to the model predominantly reflect distributional characteristics (i.e., P25, P75, and P90) in the data. Based on the computational foundation of time-series features, the explanatory variables that play a major explanatory role mainly consist of vegetation indices involving the red edge band. This was likely due to the higher sensitivity of red-edge wavelengths to vegetation growth status [86]. In savanna areas characterized by high horizontal heterogeneity, the utilization of standard deviation as the aggregation method during resampling from 10 m to 30 m resolution effectively accentuated spectral disparities between trees and grasses, thereby further enhancing the explanatory power of red-edge-based indices.
Radar features demonstrated moderate explanatory capability (Figure 5e,f), though their overall contribution remained weaker than that of optical-derived variables. While L-band radar data exhibits superior canopy penetration capacity, it potentially penetrates through the canopy to the ground surface in areas with sparse tree cover, leading to more complex backscattering features and consequently diminished explanatory power relative to C-band radar data [87]. Although estimation errors exhibited a positive correlation with slope, the predominantly flat terrain of the study area likely diminished the explanatory role of topographic features. It is important to note that the limited explanatory power observed in this study does not imply a lack of significance in modeling other regions with more varied topography. A substantial number of AI-reconstructed features were retained in the final model, yet their explanatory power showed relatively limited, possibly due to constraints caused by aggregation statistics during the resampling process. Nevertheless, the AI-reconstructed dataset provided annual composite images without clouds and cloud shadows, maintaining its value in supplying relevant explanatory information for various applications such as land cover classification and biophysical parameter retrieval [88].
This highlights the need to build models tailored to the characteristics of the savanna ecosystem, different from those designed for forest systems, based on prior ecological knowledge.

5.3. Limitation and Prospection

Despite implementing a comprehensive filtering framework on the GEDI L2A product to obtain reliable footprints, our results still exhibited overestimation in low-stature vegetation based on the RH98 metrics. This may be caused by the terrain variation in full-waveform processing. When dealing with vegetation on sloping terrain, it can be difficult to accurately extract canopy heights due to the canopy and ground echoes appearing at the same height [89]. Although GEDI has implemented a smaller footprint diameter to mitigate this issue [26], slope remains a crucial factor affecting GEDI’s relative height accuracy [80]. As the slope increases, the canopy height accuracy of GEDI decreases [90]. Previous studies have shown that canopy height prediction based on GEDI can experience a decrease in modeling accuracy when the slope increases [24,91]. However, in our study, the terrain in the KNP area is predominantly flat, with over 87% of pixels having slopes less than 6°, which minimizes the effect of slope. Nevertheless, error analysis presented a marked increase in mapping error with a rising slope (Figure 7i). Consequently, future efforts should focus on refining processing algorithms of full-waveform data to enhance the retrieval of relative height metrics.
The geolocation uncertainty of GEDI also poses a significant challenge to our study. Roy, et al. [42] assessed the impact of geolocation uncertainty on the GEDI data for forest canopy height estimation. They emphasized that the uncertainty of estimations depends on the characteristics of the forest canopy structure around the GEDI footprint locations. Therefore, caution should be advised when using GEDI data over spatially heterogeneous canopies. It should be noted that the comprehensive filtering framework employing vegetation and radar indices has undergone validation solely in the savanna area. Therefore, implementation in other global forest systems must be carefully adapted to accommodate local environmental characteristics and research requirements. Nevertheless, the approach we adopted to develop this comprehensive filtering framework can provide valuable insights for establishing similar filtering strategies in other forest systems.
The higher uncertainties exhibited by three global canopy height products in the savanna area highlight that accurate mapping in such heterogeneous landscapes requires specifically adapted model training [24,78,79]. Compared to machine learning models, the deep learning-based global products demonstrate superior performance in mitigating the overestimation of low values and underestimation of high values (Figure 7a–e). The limited availability of high-canopy GEDI samples (Figure 4b), caused by sparse tree cover and GEDI’s discrete sampling pattern, remains a primary factor affecting the precision of our localized model. Therefore, future work should focus on developing savanna-specific tree canopy height estimation models by integrating advanced deep learning algorithms with freely available multi-modal datasets, enabling more precise quantification of ecological effects [92,93].

6. Conclusions

In this study, a comprehensive filtering framework was designed to accurately extract GEDI L2A footprint data while minimizing the influence of geolocation errors on subsequent modeling. By integrating the filtered GEDI footprints with multi-source remote sensing data (i.e., Sentinel-2, Sentinel-1, PALSAR-2, SRTM, and AlphaEarth), a Random Forest regression model was constructed to estimate and produce a 30 m tree canopy height map for KNP in 2020. The predicted savanna tree canopy height showed strong agreement with the reference CHM and outperformed existing regional and global canopy height products in the study area. It accurately captured the horizontal heterogeneity of the savanna vegetation, with tall trees found predominantly along rivers and their tributaries. Analysis of influencing factors revealed that canopy cover showed a negligible impact on prediction accuracy, whereas increasing slope significantly reduced the accuracy of canopy height predictions. A tendency to overestimate low vegetation and underestimate tall vegetation was observed. Future work will focus on model refinement through the exploration of more interpretable features, effective algorithms, and improved correction of GEDI overestimation in low-stature vegetation areas. This study underscores the need to develop models specific to the complex savanna ecosystem distinct from forest CHMs. It also demonstrates the potential of integrating GEDI data with spatially continuous satellite imagery for large-scale, continuous tree canopy height mapping, thereby supporting improved monitoring and understanding of tropical savanna ecosystems.

Author Contributions

Conceptualization, X.M., Y.Q., C.X. and G.Z.; methodology, X.M.; validation, X.M., M.C. and Y.Q.; formal analysis, X.M.; investigation, X.M., Y.Q. and X.L.; resources, X.M. and Y.Q.; data curation, X.M., M.C. and Y.Q.; writing—original draft preparation, X.M. and Y.Q.; writing—review and editing, X.M., Y.Q., G.Z., C.X., M.C. and X.L.; visualization, X.M.; supervision, G.Z.; funding acquisition, G.Z. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by the National Natural Science Foundation of China (grant number: 42571464).

Data Availability Statement

Data will be made available on request.

Acknowledgments

This research was conducted at the International Institute for Earth System Science, Nanjing University. Anonymous reviewers are appreciated for their valuable suggestions and comments to improve this manuscript significantly.

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Moore, C.E.; Beringer, J.; Evans, B.; Hutley, L.B.; McHugh, I.; Tapper, N.J. The contribution of trees and grasses to productivity of an Australian tropical savanna. Biogeosciences 2016, 13, 2387–2403. [Google Scholar] [CrossRef]
  2. Scholes, R.J.; Archer, S.R. Tree-grass interactions in savannas. Annu. Rev. Ecol. Syst. 1997, 28, 517–544. [Google Scholar] [CrossRef]
  3. Bouvet, A.; Mermoz, S.; Toan, T.L.; Villard, L.; Mathieu, R.; Naidoo, L.; Asner, G.P. An above-ground biomass map of African savannahs and woodlands at 25 m resolution derived from ALOS PALSAR. Remote Sens. Environ. 2018, 206, 156–173. [Google Scholar] [CrossRef]
  4. Holdo, R.M.; Nippert, J.B. Linking resource- and disturbance-based models to explain tree-grass coexistence in savannas. New Phytol. 2023, 237, 1966–1979. [Google Scholar] [CrossRef] [PubMed]
  5. Lucas, R.M.; Lee, A.C.; Armston, J.; Carreiras, J.; Viergever, K.M.; Bunting, P.; Clewley, D.; Moghaddam, M.; Siqueira, P.; Woodhouse, I. Quantifying carbon in savannas: The role of active sensors in measurements of tree structure and biomass. In Ecosystem Function in Savannas: Measurement and Modeling at Landscape to Global Scales; Taylor & Francis: New York, NY, USA, 2010; pp. 155–174. [Google Scholar]
  6. Zhou, Y.; Singh, J.; Butnor, J.R.; Coetsee, C.; Boucher, P.B.; Case, M.F.; Hockridge, E.G.; Davies, A.B.; Staver, A.C. Limited increases in savanna carbon stocks over decades of fire suppression. Nature 2022, 603, 445–449. [Google Scholar] [CrossRef]
  7. Gaynor, K.M.; Daskin, J.H.; Rich, L.N.; Brashares, J.S. Postwar wildlife recovery in an African savanna: Evaluating patterns and drivers of species occupancy and richness. Anim. Conserv. 2021, 24, 510–522. [Google Scholar] [CrossRef]
  8. Hill, M.J.; Hanan, N.P. Ecosystem Function in Savannas: Measurement and Modeling at Landscape to Global Scales; CRC Press: Boca Raton, FL, USA, 2010. [Google Scholar]
  9. Grace, J.; Jose, J.S.; Meir, P.; Miranda, H.S.; Montes, R.A. Productivity and carbon fluxes of tropical savannas. J. Biogeogr. 2006, 33, 387–400. [Google Scholar] [CrossRef]
  10. Shackleton, C.; Shackleton, S. The importance of non-timber forest products in rural livelihood security and as safety nets: A review of evidence from South Africa. S. Afr. J. Sci. 2004, 100, 658–664. [Google Scholar]
  11. Murphy, B.P.; Bowman, D.M. What controls the distribution of tropical forest and savanna? Ecol. Lett. 2012, 15, 748–758. [Google Scholar] [CrossRef]
  12. Archer, S. Tree-grass dynamics in a Prosopis-thornscrub savanna parkland: Reconstructing the past and predicting the future. Ecoscience 1995, 2, 83–99. [Google Scholar] [CrossRef]
  13. Devine, A.P.; McDonald, R.A.; Quaife, T.; Maclean, I.M. Determinants of woody encroachment and cover in African savannas. Oecologia 2017, 183, 939–951. [Google Scholar] [CrossRef]
  14. Mitchard, E.T.; Flintrop, C.M. Woody encroachment and forest degradation in sub-Saharan Africa’s woodlands and savannas 1982–2006. Philos. Trans. R. Soc. B 2013, 368, 20120406. [Google Scholar] [CrossRef] [PubMed]
  15. Zhang, G.; Ganguly, S.; Nemani, R.R.; White, M.A.; Milesi, C.; Hashimoto, H.; Wang, W.; Saatchi, S.; Yu, Y.; Myneni, R.B. Estimation of forest aboveground biomass in California using canopy height and leaf area index estimated from satellite data. Remote Sens. Environ. 2014, 151, 44–56. [Google Scholar] [CrossRef]
  16. Wang, M.; Sun, R.; Xiao, Z. Estimation of forest canopy height and aboveground biomass from spaceborne LiDAR and Landsat imageries in Maryland. Remote Sens. 2018, 10, 344. [Google Scholar] [CrossRef]
  17. Martens, S.N.; Breshears, D.D.; Meyer, C.W. Spatial distributions of understory light along the grassland/forest continuum: Effects of cover, height, and spatial pattern of tree canopies. Ecol. Modell. 2000, 126, 79–93. [Google Scholar] [CrossRef]
  18. Pimont, F.; Dupuy, J.-L.; Linn, R.R.; Dupont, S. Impacts of tree canopy structure on wind flows and fire propagation simulated with FIRETEC. Ann. For. Sci. 2011, 68, 523–530. [Google Scholar] [CrossRef]
  19. Feng, G.; Zhang, J.; Girardello, M.; Pellissier, V.; Svenning, J.C. Forest canopy height co-determines taxonomic and functional richness, but not functional dispersion of mammals and birds globally. Glob. Ecol. Biogeogr. 2020, 29, 1350–1359. [Google Scholar] [CrossRef]
  20. Davies, A.B.; Asner, G.P. Advances in animal ecology from 3D-LiDAR ecosystem mapping. Trends Ecol. Evol. 2014, 29, 681–691. [Google Scholar] [CrossRef] [PubMed]
  21. Muumbe, T.P.; Baade, J.; Singh, J.; Schmullius, C.; Thau, C. Terrestrial laser scanning for vegetation analyses with a special focus on savannas. Remote Sens. 2021, 13, 507. [Google Scholar] [CrossRef]
  22. Boucher, P.B.; Hockridge, E.G.; Singh, J.; Davies, A.B. Flying high: Sampling savanna vegetation with UAV-lidar. Methods Ecol. Evol. 2023, 14, 1668–1686. [Google Scholar] [CrossRef]
  23. Heckel, K.; Urban, M.; Bouffard, J.-S.; Baade, J.; Boucher, P.; Davies, A.; Hockridge, E.G.; Luck, W.; Ziemer, J.; Smit, I.; et al. The first sub-meter resolution digital elevation model of the Kruger National Park, South Africa. Koedoe 2021, 63, 13. [Google Scholar] [CrossRef]
  24. Potapov, P.; Li, X.; Hernandez-Serna, A.; Tyukavina, A.; Hansen, M.C.; Kommareddy, A.; Pickens, A.; Turubanova, S.; Tang, H.; Silva, C.E. Mapping global forest canopy height through integration of GEDI and Landsat data. Remote Sens. Environ. 2021, 253, 112165. [Google Scholar] [CrossRef]
  25. Wang, S.; Liu, C.; Li, W.; Jia, S.; Yue, H. Hybrid model for estimating forest canopy heights using fused multimodal spaceborne LiDAR data and optical imagery. Int. J. Appl. Earth Obs. Geoinf. 2023, 122, 103431. [Google Scholar] [CrossRef]
  26. Dubayah, R.; Blair, J.B.; Goetz, S.; Fatoyinbo, L.; Hansen, M.; Healey, S.; Hofton, M.; Hurtt, G.; Kellner, J.; Luthcke, S. The Global Ecosystem Dynamics Investigation: High-resolution laser ranging of the Earth’s forests and topography. Sci. Remote Sens. 2020, 1, 100002. [Google Scholar] [CrossRef]
  27. Li, X.X.; Wessels, K.; Armston, J.; Hancock, S.; Mathieu, R.; Main, R.; Naidoo, L.; Erasmus, B.; Scholes, R. First validation of GEDI canopy heights in African savannas. Remote Sens. Environ. 2023, 285, 113402. [Google Scholar] [CrossRef]
  28. García, M.; Saatchi, S.; Ustin, S.; Balzter, H. Modelling forest canopy height by integrating airborne LiDAR samples with satellite Radar and multispectral imagery. Int. J. Appl. Earth Obs. Geoinf. 2018, 66, 159–173. [Google Scholar] [CrossRef]
  29. Fagua, J.C.; Jantz, P.; Rodriguez-Buritica, S.; Duncanson, L.; Goetz, S.J. Integrating LiDAR, multispectral and SAR data to estimate and map canopy height in tropical forests. Remote Sens. 2019, 11, 2697. [Google Scholar] [CrossRef]
  30. de Almeida, C.T.; Gerente, J.; Campos, J.R.D.; Gomes, F.C.; Providelo, L.A.; Marchiori, G.; Chen, X.J. Canopy Height Mapping by Sentinel 1 and 2 Satellite Images, Airborne LiDAR Data, and Machine Learning. Remote Sens. 2022, 14, 4112. [Google Scholar] [CrossRef]
  31. Pourshamsi, M.; Xia, J.S.; Yokoya, N.; Garcia, M.; Lavalle, M.; Pottier, E.; Balzter, H. Tropical forest canopy height estimation from combined polarimetric SAR and LiDAR using machine-learning. ISPRS J. Photogramm. Remote Sens. 2021, 172, 79–94. [Google Scholar] [CrossRef]
  32. Jiang, F.; Zhao, F.; Ma, K.; Li, D.; Sun, H. Mapping the forest canopy height in Northern China by synergizing ICESat-2 with Sentinel-2 using a stacking algorithm. Remote Sens. 2021, 13, 1535. [Google Scholar] [CrossRef]
  33. Potapov, P.; Tyukavina, A.; Turubanova, S.; Talero, Y.; Hernandez-Serna, A.; Hansen, M.; Saah, D.; Tenneson, K.; Poortinga, A.; Aekakkararungroj, A. Annual continuous fields of woody vegetation structure in the Lower Mekong region from 2000-2017 Landsat time-series. Remote Sens. Environ. 2019, 232, 111278. [Google Scholar] [CrossRef]
  34. Campbell, M.J.; Dennison, P.E.; Kerr, K.L.; Brewer, S.C.; Anderegg, W.R. Scaled biomass estimation in woodland ecosystems: Testing the individual and combined capacities of satellite multispectral and lidar data. Remote Sens. Environ. 2021, 262, 112511. [Google Scholar] [CrossRef]
  35. Li, W.; Niu, Z.; Shang, R.; Qin, Y.; Wang, L.; Chen, H. High-resolution mapping of forest canopy height using machine learning by coupling ICESat-2 LiDAR with Sentinel-1, Sentinel-2 and Landsat-8 data. Int. J. Appl. Earth Obs. Geoinf. 2020, 92, 102163. [Google Scholar] [CrossRef]
  36. Xi, Z.; Xu, H.; Xing, Y.; Gong, W.; Chen, G.; Yang, S. Forest canopy height mapping by synergizing ICESat-2, Sentinel-1, Sentinel-2 and topographic information based on machine learning methods. Remote Sens. 2022, 14, 364. [Google Scholar] [CrossRef]
  37. Brown, C.F.; Kazmierski, M.R.; Pasquarella, V.J.; Rucklidge, W.J.; Samsikova, M.; Zhang, C.; Shelhamer, E.; Lahera, E.; Wiles, O.; Ilyushchenko, S. AlphaEarth Foundations: An embedding field model for accurate and efficient global mapping from sparse label data. arXiv 2025, arXiv:2507.22291. [Google Scholar] [CrossRef]
  38. Tang, H.; Stoker, J.; Luthcke, S.; Armston, J.; Lee, K.; Blair, B.; Hofton, M. Evaluating and mitigating the impact of systematic geolocation error on canopy height measurement performance of GEDI. Remote Sens. Environ. 2023, 291, 113571. [Google Scholar] [CrossRef]
  39. Ngo, Y.-N.; Ho Tong Minh, D.; Baghdadi, N.; Fayad, I. Tropical Forest Top Height by GEDI: From Sparse Coverage to Continuous Data. Remote Sens. 2023, 15, 975. [Google Scholar] [CrossRef]
  40. Schleich, A.; Durrieu, S.; Vega, C. Improving GEDI Footprint Geolocation Using a High Resolution Digital Elevation Model. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2023, 16, 7718–7732. [Google Scholar] [CrossRef]
  41. Moudrý, V.; Prošek, J.; Marselis, S.; Marešová, J.; Šárovcová, E.; Gdulová, K.; Kozhoridze, G.; Torresani, M.; Rocchini, D.; Eltner, A. How to Find Accurate Terrain and Canopy Height GEDI Footprints in Temperate Forests and Grasslands? Earth Space Sci. 2024, 11, e2024EA003709. [Google Scholar] [CrossRef]
  42. Roy, D.P.; Kashongwe, H.B.; Armston, J. The impact of geolocation uncertainty on GEDI tropical forest canopy height estimation and change monitoring. Sci. Remote Sens. 2021, 4, 100024. [Google Scholar] [CrossRef]
  43. Yang, C.; Peng, D.; Deng, K.; Jiang, L.; Zhao, M.; Zeng, W.; Shao, Y.; Wang, N. Quantification of GEDI Geolocation Error and its Influence on Elevation and Canopy Height. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2024, 17, 18012–18024. [Google Scholar] [CrossRef]
  44. Bunting, E.L.; Fullman, T.; Kiker, G.; Southworth, J. Utilization of the SAVANNA model to analyze future patterns of vegetation cover in Kruger National Park under changing climate. Ecol. Modell. 2016, 342, 147–160. [Google Scholar] [CrossRef]
  45. Rountree, M.W.; Rogers, K.H.; Heritage, G.L. LANDSCAPE STATE CHANGE IN THE SEMI-ARID SABIE RIVER, KRUGER NATIONAL PARK, IN RESPONSE TO FLOOD AND DROUGHT. S. Afr. Geogr. J. 2000, 82, 173–181. [Google Scholar] [CrossRef]
  46. MacFadyen, S.; Zambatis, N.; Van Teeffelen, A.J.A.; Hui, C. Long-term rainfall regression surfaces for the Kruger National Park, South Africa: A spatio-temporal review of patterns from 1981 to 2015. Int. J. Climatol. 2018, 38, 2506–2519. [Google Scholar] [CrossRef]
  47. Ben-Shahar, R. Does fencing reduce the carrying capacity for populations of large herbivores? J. Trop. Ecol. 1993, 9, 249–253. [Google Scholar] [CrossRef]
  48. Venter, F.; Gertenbach, W. A cursory review of the climate and vegetation of the Kruger National Park. Koedoe 1986, 29, 139–148. [Google Scholar] [CrossRef]
  49. Eckhardt, H.; Van Wilgen, B.; Biggs, H. Trends in woody vegetation cover in the Kruger National Park, South Africa, between 1940 and 1998. Afr. J. Ecol. 2000, 38, 108–115. [Google Scholar] [CrossRef]
  50. Beck, J.; Wirt, B.; Armston, J.; Hofton, M.; Luthcke, S.; Tang, H. Global Ecosystem Dynamics Investigation (GEDI) Level 02 User Guide; NASA: Washington, DC, USA, 2021; Volume 2.
  51. Hofton, M.; Blair, J.B.; Story, S.; Yi, D. Algorithm Theoretical Basis Document (ATBD); NASA: Washington, DC, USA, 2020.
  52. Blair, J.B.; Hofton, M.A. Modeling laser altimeter return waveforms over complex vegetation using high-resolution elevation data. Geophys. Res. Lett. 1999, 26, 2509–2512. [Google Scholar] [CrossRef]
  53. Drusch, M.; Del Bello, U.; Carlier, S.; Colin, O.; Fernandez, V.; Gascon, F.; Hoersch, B.; Isola, C.; Laberinti, P.; Martimort, P. Sentinel-2: ESA’s optical high-resolution mission for GMES operational services. Remote Sens. Environ. 2012, 120, 25–36. [Google Scholar] [CrossRef]
  54. Phiri, D.; Simwanda, M.; Salekin, S.; Nyirenda, V.R.; Murayama, Y.; Ranagalage, M. Sentinel-2 Data for Land Cover/Use Mapping: A Review. Remote Sens. 2020, 12, 2291. [Google Scholar] [CrossRef]
  55. Rouse, J. Monitoring the Vernal Advancement and Retrogradation of Natural Vegetation; NASA: Washington, DC, USA, 1973.
  56. Qi, J.; Chehbouni, A.; Huete, A.R.; Kerr, Y.H.; Sorooshian, S. A modified soil adjusted vegetation index. Remote Sens. Environ. 1994, 48, 119–126. [Google Scholar] [CrossRef]
  57. Gillespie, A.R.; Kahle, A.B.; Walker, R.E. Color enhancement of highly correlated images. II. Channel ratio and “chromaticity” transformation techniques. Remote Sens. Environ. 1987, 22, 343–365. [Google Scholar] [CrossRef]
  58. Woebbecke, D.M.; Meyer, G.E.; Von Bargen, K.; Mortensen, D.A. Color indices for weed identification under various soil, residue, and lighting conditions. Trans. ASAE 1995, 38, 259–269. [Google Scholar] [CrossRef]
  59. Drusch, M.; Gascon, F.; Berger, M. GMES Sentinel-2 Mission Requirements Document; European Space Agency: Paris, France, 2010. [Google Scholar]
  60. Fitzgerald, G.; Rodriguez, D.; O’Leary, G. Measuring and predicting canopy nitrogen nutrition in wheat using a spectral index—The canopy chlorophyll content index (CCCI). Field Crops Res. 2010, 116, 318–324. [Google Scholar] [CrossRef]
  61. Dash, J.; Curran, P.J. The MERIS terrestrial chlorophyll index. Int. J. Remote Sens. 2004, 25, 5403–5413. [Google Scholar] [CrossRef]
  62. Frampton, W.J.; Dash, J.; Watmough, G.; Milton, E.J. Evaluating the capabilities of Sentinel-2 for quantitative estimation of biophysical variables in vegetation. ISPRS J. Photogramm. Remote Sens. 2013, 82, 83–92. [Google Scholar] [CrossRef]
  63. Haboudane, D.; Miller, J.R.; Tremblay, N.; Zarco-Tejada, P.J.; Dextraze, L. Integrated narrow-band vegetation indices for prediction of crop chlorophyll content for application to precision agriculture. Remote Sens. Environ. 2002, 81, 416–426. [Google Scholar] [CrossRef]
  64. Rosenqvist, A.; Killough, B. A Layman’s Interpretation Guide to L-Band and C-Band Synthetic Aperture Radar Data; Comittee on Earth Observation Satellites: Washington, DC, USA, 2018. [Google Scholar]
  65. Atmani, F.; Bookhagen, B.; Smith, T. Measuring Vegetation Heights and Their Seasonal Changes in the Western Namibian Savanna Using Spaceborne Lidars. Remote Sens. 2022, 14, 2928. [Google Scholar] [CrossRef]
  66. Farr, T.G.; Rosen, P.A.; Caro, E.; Crippen, R.; Duren, R.; Hensley, S.; Kobrick, M.; Paller, M.; Rodriguez, E.; Roth, L. The shuttle radar topography mission. Rev. Geophys. 2007, 45, RG2004. [Google Scholar] [CrossRef]
  67. Jucker, T.; Bongalov, B.; Burslem, D.; Nilus, R.; Dalponte, M.; Lewis, S.L.; Phillips, O.L.; Qie, L.; Coomes, D.A. Topography shapes the structure, composition and function of tropical forest landscapes. Ecol. Lett. 2018, 21, 989–1000. [Google Scholar] [CrossRef] [PubMed]
  68. Reiner, F.; Brandt, M.; Tong, X.; Skole, D.; Kariryaa, A.; Ciais, P.; Davies, A.; Hiernaux, P.; Chave, J.; Mugabowindekwe, M.; et al. Africa Tree Cover Map; Zenodo: Geneva, Switzerland, 2023. [Google Scholar] [CrossRef]
  69. Reiner, F.; Brandt, M.; Tong, X.; Skole, D.; Kariryaa, A.; Ciais, P.; Davies, A.; Hiernaux, P.; Chave, J.; Mugabowindekwe, M. More than one quarter of Africa’s tree cover is found outside areas previously classified as forest. Nat. Commun. 2023, 14, 2258. [Google Scholar] [CrossRef]
  70. Lagomasino, D.; Fatoyinbo, T.; Lee, S.-K.; Simard, M. High-resolution forest canopy height estimation in an African blue carbon ecosystem. Remote Sens. Ecol. Conserv. 2015, 1, 51–60. [Google Scholar] [CrossRef]
  71. Filippelli, S.K.; Vogeler, J.C.; Mauro, F.; Coetsee, C.; Fekety, P.A.; McHale, M.; Bunn, D. Tracking savanna vegetation structure in South Africa by extension of GEDI canopy metrics with Landsat, Sentinel-2, and PALSAR. Sci. Remote Sens. 2026, 13, 100388. [Google Scholar] [CrossRef]
  72. Chen, P.; Huang, H.; Qin, P.; Wu, Z.; Wang, Z.; Liu, C.; Dong, N.; Wang, J. Unveiling the performance and influential factors of GEDI L2A for building height retrieval. GIScience Remote Sens. 2025, 62, 2498785. [Google Scholar] [CrossRef]
  73. Wright, M.N.; Ziegler, A. ranger: A Fast Implementation of Random Forests for High Dimensional Data in C++ and R. J. Stat. Softw. 2017, 77, 1–17. [Google Scholar] [CrossRef]
  74. Breiman, L. Random forests. Mach. Learn. 2001, 45, 5–32. [Google Scholar] [CrossRef]
  75. Jin, S.; Su, Y.; Gao, S.; Hu, T.; Liu, J.; Guo, Q. The transferability of Random Forest in canopy height estimation from multi-source remote sensing data. Remote Sens. 2018, 10, 1183. [Google Scholar] [CrossRef]
  76. Ahmed, O.S.; Franklin, S.E.; Wulder, M.A.; White, J.C. Characterizing stand-level forest canopy cover and height using Landsat time series, samples of airborne LiDAR, and the Random Forest algorithm. ISPRS J. Photogramm. Remote Sens. 2015, 101, 89–101. [Google Scholar] [CrossRef]
  77. Olden, J.D.; Lawler, J.J.; Poff, N.L. Machine learning methods without tears: A primer for ecologists. Q. Rev. Biol. 2008, 83, 171–193. [Google Scholar] [CrossRef]
  78. Lang, N.; Jetz, W.; Schindler, K.; Wegner, J.D. A high-resolution canopy height model of the Earth. Nat. Ecol. Evol. 2023, 7, 1778–1789. [Google Scholar] [CrossRef]
  79. Tolan, J.; Yang, H.-I.; Nosarzewski, B.; Couairon, G.; Vo, H.V.; Brandt, J.; Spore, J.; Majumdar, S.; Haziza, D.; Vamaraju, J. Very high resolution canopy height maps from RGB imagery using self-supervised vision transformer and convolutional decoder trained on aerial lidar. Remote Sens. Environ. 2024, 300, 113888. [Google Scholar] [CrossRef]
  80. Wang, C.; Elmore, A.J.; Numata, I.; Cochrane, M.A.; Shaogang, L.; Huang, J.; Zhao, Y.; Li, Y. Factors affecting relative height and ground elevation estimations of GEDI among forest types across the conterminous USA. GIScience Remote Sens. 2022, 59, 975–999. [Google Scholar] [CrossRef]
  81. Kutchartt, E.; Pedron, M.; Pirotti, F. Assessment of canopy and ground height accuracy from gedi lidar over steep mountain areas. ISPRS Ann. Photogramm. Remote Sens. Spat. Inf. Sci. 2022, 3, 431–438. [Google Scholar] [CrossRef]
  82. Huang, J.; Xia, T.; Shuai, Y.; Zhu, H. Assessing the Performance of GEDI LiDAR Data for Estimating Terrain in Densely Forested Areas. IEEE Geosci. Remote Sens. Lett. 2023, 20, 1–5. [Google Scholar] [CrossRef]
  83. Higgins, S.I.; Delgado-Cartay, M.D.; February, E.C.; Combrink, H.J. Is there a temporal niche separation in the leaf phenology of savanna trees and grasses? J. Biogeogr. 2011, 38, 2165–2175. [Google Scholar] [CrossRef]
  84. Archibald, S.; Scholes, R. Leaf green-up in a semi-arid African savanna-separating tree and grass responses to environmental cues. J. Veg. Sci. 2007, 18, 583–594. [Google Scholar]
  85. Luo, Y.; El-Madany, T.S.; Filippa, G.; Ma, X.; Ahrens, B.; Carrara, A.; Gonzalez-Cascon, R.; Cremonese, E.; Galvagno, M.; Hammer, T.W. Using near-infrared-enabled digital repeat photography to track structural and physiological phenology in Mediterranean tree–grass ecosystems. Remote Sens. 2018, 10, 1293. [Google Scholar] [CrossRef]
  86. Li, D.; Chen, J.M.; Yu, W.; Zheng, H.; Yao, X.; Zhu, Y.; Cao, W.; Cheng, T. A chlorophyll-constrained semi-empirical model for estimating leaf area index using a red-edge vegetation index. Comput. Electron. Agric. 2024, 220, 108891. [Google Scholar] [CrossRef]
  87. Viergever, K.M. Establishing the Sensitivity of Synthetic Aperture Radar to Above-Ground Biomass in Wooded Savannas. Ph.D. Thesis, The University of Edinburgh, Edinburgh, UK, 2008. [Google Scholar]
  88. Zhang, C.; Kerner, H.; Wang, S.; Hao, P.; Li, Z.; Hunt, K.A.; Abernethy, J.; Zhao, H.; Gao, F.; Di, L.; et al. Remote sensing for crop mapping: A perspective on current and future crop-specific land cover data products. Remote Sens. Environ. 2025, 330, 114995. [Google Scholar] [CrossRef]
  89. Lefsky, M.A.; Keller, M.; Pang, Y.; De Camargo, P.B.; Hunter, M.O. Revised method for forest canopy height estimation from Geoscience Laser Altimeter System waveforms. J. Appl. Remote Sens. 2007, 1, 013537. [Google Scholar]
  90. Adam, M.; Urbazaev, M.; Dubois, C.; Schmullius, C. Accuracy assessment of GEDI terrain elevation and canopy height estimates in European temperate forests: Influence of environmental and acquisition parameters. Remote Sens. 2020, 12, 3948. [Google Scholar] [CrossRef]
  91. Simard, M.; Pinto, N.; Fisher, J.B.; Baccini, A. Mapping forest canopy height globally with spaceborne lidar. J. Geophys. Res. Biogeosci. 2011, 116, 4021. [Google Scholar] [CrossRef]
  92. Su, Y.; Schwartz, M.; Fayad, I.; García, M.; Zavala, M.A.; Tijerín-Triviño, J.; Astigarraga, J.; Cruz-Alonso, V.; Liu, S.; Zhang, X.; et al. Canopy height and biomass distribution across the forests of Iberian Peninsula. Sci. Data 2025, 12, 678. [Google Scholar] [CrossRef] [PubMed]
  93. Dixon, D.J.; Zhu, Y.; Jin, Y. Canopy height estimation from PlanetScope time series with spatio-temporal deep learning. Remote Sens. Environ. 2025, 318, 114518. [Google Scholar] [CrossRef]
Remotesensing 18 01523 g001
Figure 1. Study area and data preparation: location of Kruger National Park (KNP) in South Africa (a); tree canopy cover (a,d); tree canopy height model (CHM) for 2018 (b,e); filtered GEDI footprints for 2020 (c,f); corresponding close-up views (d–f); and available observations from Sentinel-2 (g), Sentinel-1 (h), and PALSAR-2 (i) for 2020.
Figure 1. Study area and data preparation: location of Kruger National Park (KNP) in South Africa (a); tree canopy cover (a,d); tree canopy height model (CHM) for 2018 (b,e); filtered GEDI footprints for 2020 (c,f); corresponding close-up views (d–f); and available observations from Sentinel-2 (g), Sentinel-1 (h), and PALSAR-2 (i) for 2020.
Remotesensing 18 01523 g001
Remotesensing 18 01523 g002
Figure 2. The flowchart and main components of tree canopy height mapping in this study.
Figure 2. The flowchart and main components of tree canopy height mapping in this study.
Remotesensing 18 01523 g002
Remotesensing 18 01523 g003
Figure 3. Accuracy assessment of the filtered GEDI footprints by simplified (a) and comprehensive filtering framework (b) methods, respectively.
Figure 3. Accuracy assessment of the filtered GEDI footprints by simplified (a) and comprehensive filtering framework (b) methods, respectively.
Remotesensing 18 01523 g003
Remotesensing 18 01523 g004
Figure 4. Frequency distribution of footprints from simplified (a) and comprehensive filtering framework (b) methods, with and without sampling; the accuracy assessment of estimated models by Pearson’s r (c) and RMSE (d) metrics; scatter density plots of predictions versus reference CHM data (e,f); and RMSE values across different tree cover levels (g).
Figure 4. Frequency distribution of footprints from simplified (a) and comprehensive filtering framework (b) methods, with and without sampling; the accuracy assessment of estimated models by Pearson’s r (c) and RMSE (d) metrics; scatter density plots of predictions versus reference CHM data (e,f); and RMSE values across different tree cover levels (g).
Remotesensing 18 01523 g004
Remotesensing 18 01523 g005
Figure 5. Feature importance of the retained features derived from AlphaEarth (a), Sentinel-2 (b–d), Sentinel-1 (e), and PALSAR-2 (f) data used for training models.
Figure 5. Feature importance of the retained features derived from AlphaEarth (a), Sentinel-2 (b–d), Sentinel-1 (e), and PALSAR-2 (f) data used for training models.
Remotesensing 18 01523 g005
Remotesensing 18 01523 g006
Figure 6. The tree canopy height maps of KNP with a spatial resolution of 30 m (a) and four close-up views of tree canopy height at pixel level (b(1I,2I,3I,4I)) were compared with the reference CHM data (b(1II,2II,3II,4II)) and very high-resolution imagery (b(1III,2III,3III,4III)).
Figure 6. The tree canopy height maps of KNP with a spatial resolution of 30 m (a) and four close-up views of tree canopy height at pixel level (b(1I,2I,3I,4I)) were compared with the reference CHM data (b(1II,2II,3II,4II)) and very high-resolution imagery (b(1III,2III,3III,4III)).
Remotesensing 18 01523 g006
Remotesensing 18 01523 g007
Figure 7. Accuracy assessment of existing products (a–d) [24,71,78,79] and our results (e); distribution of residuals across all mapping results (f); relationships between estimation errors and key impact factors: tree canopy height (g), tree cover (h), and slope (i).
Figure 7. Accuracy assessment of existing products (a–d) [24,71,78,79] and our results (e); distribution of residuals across all mapping results (f); relationships between estimation errors and key impact factors: tree canopy height (g), tree cover (h), and slope (i).
Remotesensing 18 01523 g007
Remotesensing 18 01523 g008
Figure 8. Close-up comparison of our mapping results (a1–d1) against existing products and reference data: Filippelli et al. [71] (a2–d2); Lang et al. [78] (a3–d3); Potapov et al. [24] (a4–d4); and Tolan et al. [79] (a5–d5); the reference CHM (a6–d6); and very high-resolution imagery (a7–d7).
Figure 8. Close-up comparison of our mapping results (a1–d1) against existing products and reference data: Filippelli et al. [71] (a2–d2); Lang et al. [78] (a3–d3); Potapov et al. [24] (a4–d4); and Tolan et al. [79] (a5–d5); the reference CHM (a6–d6); and very high-resolution imagery (a7–d7).
Remotesensing 18 01523 g008
Table 1. All features derived from different data sources.
Table 1. All features derived from different data sources.
Data SourcesRaw BandsFeaturesMultitemporal StatisticsAggregation
(Mean/Std)		Number
Sentinel-2
(2020.01–2020.12)		B2(Blue)
B3(Green)
B4(Red)
B5(Red Edge1)
B6(Red Edge2)
B7(Red Edge3)
B8(NIR)
B8A(Narrow NIR)
B11(SWIR1)
B12(SWIR2)		NDVI_B84 = (B8 − B4)/(B8 + B4)
NDVI_B85 = (B8 − B5)/(B8 + B5)
NDVI_B86 = (B8 − B6)/(B8 + B6)
NDVI_B87 = (B8 − B7)/(B8 + B7)
NDVI_B8A4 = (B8 − B4)/(B8 + B4)
NDVI_B8A5 = (B8 − B5)/(B8 + B5)
NDVI_B8A6 = (B8 − B6)/(B8 + B6)
NDVI_B8A7 = (B8 − B7)/(B8 + B7)
S2REP = 705 + 35((((B7 + B4)/2) − B5)/(B6 − B5))
TCARI = 3((B5 − B4) − 0.2(B5 − B3)(B5/B4))
MTCI = (RE2 − RE1)/(RE1 − R)
CCCI = ((B8 − B5)/(B8 + B5))/((B8 − B4)/(B8 + B4))
GLI = (2B3 − B4 − B2)/(2B3 + B4 + B2)
ExG = 2B3 − (B4 + B2)
GCC= B3/(B4 + B3 + B2)
BCC = B2/(B4 + B3 + B2)
MSAVI = ( 2 B 8 + 1 − 2 B 8 + 1 2 − 8 ( B 8 − B 4 ) ) / 2
EVI = 2.5(B8 − B4)/(B8 + 6B4 + 7.5B2 + 1)		Mean
Variance
P10
P25
P75
P90
P100–P75
P75–P25
P25–P0		√504
Sentinel-1
(2020.01–2020.12)		VV
VH		VHVV = VH/VV√54
PALSAR-2
(2020.01–2020.12)		HH
HV		HHHV = HH/HV-27
AlphaEarth
(2020)		A00~A63--√128
SRTMElevationAspect/Slope--3
Table 2. Validation Results.
Table 2. Validation Results.
Validated by CHMValidated by RH98
Pearson’s rRMSE (m)Pearson’s rRMSE (m)
GEDI RH98Simplified0.434.10--
Comprehensive0.513.88--
Model performanceSimplified0.673.900.702.31
Comprehensive0.693.780.722.03
Filippelli et al. [71]All0.155.62--
Low [2.35, 10)0.042.83--
Medium [10, 20)0.107.61--
High [≥20, max)0.2114.59--
Lang et al. [78]All0.455.22--
Low [2.35, 10)0.134.76--
Medium [10, 20)0.345.76--
High [≥20, max)0.377.25--
Potapov et al. [24]All0.196.91--
Low [2.35, 10)0.082.95--
Medium [10, 20)0.099.83--
High [≥20, max)0.2417.03--
Tolan et al. [79]All0.444.98--
Low [2.35, 10)0.213.95--
Medium [10, 20)0.265.79--
High [≥20, max)0.0910.60--
Our ResultsAll0.664.09--
Low [2.35, 10)0.284.28--
Medium [10, 20)0.493.17--
High [≥20, max)0.287.83--
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Ma, X.; Qu, Y.; Chen, M.; Zheng, G.; Xu, C.; Li, X. Characterizing Savanna Tree Canopy Heights Using GEDI and Spatially Continuous Multi-Source Data at a Landscape Level. Remote Sens. 2026, 18, 1523. https://doi.org/10.3390/rs18101523

AMA Style

Ma X, Qu Y, Chen M, Zheng G, Xu C, Li X. Characterizing Savanna Tree Canopy Heights Using GEDI and Spatially Continuous Multi-Source Data at a Landscape Level. Remote Sensing. 2026; 18(10):1523. https://doi.org/10.3390/rs18101523

Chicago/Turabian Style

Ma, Xiao, Yajie Qu, Meiyuan Chen, Guang Zheng, Chi Xu, and Xiaoxuan Li. 2026. "Characterizing Savanna Tree Canopy Heights Using GEDI and Spatially Continuous Multi-Source Data at a Landscape Level" Remote Sensing 18, no. 10: 1523. https://doi.org/10.3390/rs18101523

APA Style

Ma, X., Qu, Y., Chen, M., Zheng, G., Xu, C., & Li, X. (2026). Characterizing Savanna Tree Canopy Heights Using GEDI and Spatially Continuous Multi-Source Data at a Landscape Level. Remote Sensing, 18(10), 1523. https://doi.org/10.3390/rs18101523

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop