Located in south-central Africa, Zambia spans 752,618 square kilometres of diverse terrain. The landscape averages 1200 meters in elevation, with vegetation blanketing nearly two-thirds of the country. Its tropical Savanna climate features three pronounced seasonal variations, where yearly precipitation shifts from 700 mm in lower elevations to over 1400 mm in northern zones. These altitudinal gradients support distinct ecosystems—from the lush Miombo woodlands dominating northern areas to the open grasslands characteristic of southern and eastern regions ( Figure 1 , Table 1 ).

This research combined three key remote sensing datasets to estimate canopy height across Zambia. The datasets used include NASA’s GEDI L2A waveform LiDAR ( Dubayah et al., 2020 ), ESA’s Sentinel-1 C-band SAR ( Torres et al., 2012 ), and Sentinel-2 multispectral imagery ( Claverie et al., 2018 ), all obtained via Google Earth Engine for the year 2020. Additional elevation data from the Shuttle Radar Topography Mission (SRTM DEM) ( Farr et al., 2007 ) was included to adjust for topographic effects. The dataset was partitioned into training (70%) and validation (30%) subsets through stratified random sampling.

This research combined three key remote sensing datasets to estimate canopy height across Zambia. The datasets used include NASA’s GEDI L2A waveform LiDAR ( Dubayah et al., 2020 ), ESA’s Sentinel-1 C-band SAR ( Torres et al., 2012 ), and Sentinel-2 multispectral imagery ( Claverie et al., 2018 ), all obtained via Google Earth Engine for the year 2020. Additional elevation data from the Shuttle Radar Topography Mission (SRTM DEM) ( Farr et al., 2007 ) was included to adjust for topographic effects. The dataset was partitioned into training (70%) and validation (30%) subsets through stratified random sampling.

The Sentinel-1 GRD imagery (VV/VH polarizations) underwent preprocessing that included eliminating thermal noise, performing radiometric calibration, and applying terrain corrections with SRTM DEM data ( Qi et al., 2019 ). To minimize speckle interference without compromising structural features critical for canopy height analysis, we implemented a Gamma-MAP filter following Kupidura (2016) methodology.

The S1 data includes images that have been taken using dual polarization (VV and VH) and in the interferometric wide swath (IW) mode. This selection was limited to only contain the ascending orbit passes and hence reducing the variability in backscatter data and ensuring uniformity in the seeing geometry. “Sentinel-1 mosaic is very suitable for monitoring forest phenology, we used it to calculate VV/VH ratio ( Soudani et al., 2021 ).”

Sentinel-2’s multispectral data (10 - 60 m resolution) contributed 22% of model predictive power through its near-infrared (NIR, band 8) and red-edge bands (bands 5-7). We used Level-2A surface reflectance products from the harmonized Landsat-Sentinel dataset ( Claverie et al., 2018 ), with cloud masking via the Scene Classification Layer (SCL). The normalized difference vegetation index (NDVI) and enhanced vegetation index (EVI) derived from these bands proved critical for discriminating leaf area in savannas (6.2 m height class) and detecting canopy gaps ( Shoko & Mutanga, 2017 ; Forkuor et al., 2017 ) ( Table 2 ).

Twelve monthly composites (2020) were generated to mitigate cloud cover, with dry-season (May-October) imagery weighted higher due to stable phenological conditions (Mutanga et al., 2012). The red-edge bands (705 - 783 nm) improved separation of grassland (2.7 - 6.7 m) and woodland canopies by detecting subtle chlorophyll content variations ( Broge & Leblanc, 2001 ). We resampled all bands to 20 m using bilinear interpolation to match GEDI’s footprint scale, preserving spectral fidelity ( Gómez, 2017 ).

The multi-sensor dataset underwent rigorous preprocessing to ensure geometric and radiometric consistency. GEDI LiDAR waveforms were filtered (sensitivity > 0.95) and normalized using a 30 m SRTM DEM to correct for topographic effects. Sentinel-2 Level-2A surface reflectance data were cloud-masked using Fmask 4.0 and aggregated into monthly composites (2020), with emphasis on dry-season acquisitions for optimal vegetation signal ( Zhu & Woodcock, 2012 ). Sentinel-1 GRD images were terrain-corrected and processed to generate VH/VV backscatter and texture features at 20 m resolution ( Vafaei et al., 2018 ). All datasets were co-registered to UTM Zone 35S (WGS84) in Google Earth Engine ( Tamiminia et al., 2020 ), with spectral indices (NDVI, EVI) calculated from Sentinel-2 and temporal alignment enforced (±15 days between GEDI and optical/SAR acquisitions). The final feature set included 24,530 quality-controlled samples stratified by biome, with 70% allocated for model training and 30% for validation ( Ahmed et al., 2015 ).

The RF algorithm’s strength, simplicity, and ability to handle complex datasets make it highly useful in various domains, including forestry, remote sensing, and environmental modeling. RF, an ensemble learning method, builds multiple decision trees during training and averages their outputs to prevent overfitting and improve model stability ( Breiman, 2001 ). Given the significant variability and non-linear relationships in data, RF is particularly effective for tasks such as estimating canopy height.

Beyond its predictive capabilities, RF’s built-in feature importance analysis enables researchers to pinpoint key explanatory variables, facilitating both dataset refinement and more insightful interpretation of modeling outcomes. However, RF is computationally intensive for large datasets and may require substantial processing power when applied to national or global scales, such as modeling forest attributes across Zambia or other countries.

Despite its many advantages, RF models can be prone to overfitting if parameters such as the number of trees, maximum depth, and minimum sample splits are not carefully tuned. Additionally, RF models are non-parametric and inherently opaque, which can limit their interpretability compared to simpler models.

The RF algorithm, introduced by Breiman (2001) , builds an ensemble of decision trees to enhance model stability and reduce overfitting. Its ability to evaluate feature importance has made it widely applicable in forest modeling. As suggested by Rodriguez-Galiano et al. (2012) , careful tuning of hyperparameters such as tree depth and sample size is critical. In our study, we set the number of trees at 50, maximum depth at 12, and implemented a 70/30 training-testing split to balance computational cost with performance.

The Random Forest model combining GEDI, Sentinel-1/2, and topographic data achieved strong accuracy (R² = 0.81, RMSE = 1.94 m training; R² = 0.78, RMSE = 2.15 m validation). This confirms its reliability for Zambia-wide canopy height estimation.

Canopy heights varied significantly by ecosystem ( Figure 2 ). Woody savannas showed the tallest (8 - 25 m) and most variable canopies, while grasslands were shortest (2.5 - 5.5 m). Savannas exhibited intermediate heights (~11 m), with combined data revealing Zambia’s structural diversity (peak at 7 m, tail > 14 m).

Forested areas (woody/savannas) displayed complex vertical structure versus simpler grasslands. This validates the model’s ability to distinguish ecosystem types using height thresholds.

Precipitation showed the strongest positive correlation with height (r = 0.56), especially in woody savannas. Temperature had weak negative effects (−0.31 in grasslands), while solar radiation was moderately positive in woody areas (r = 0.43) but less influential elsewhere.

1) Current validation data may not fully represent Zambia’s diversity—future studies should add targeted field plots and UAV-LiDAR.

2) Uncertainty propagation needs deeper analysis using methods like Monte Carlo simulations.

3) Sensor-specific contributions to error require quantification via variance decomposition techniques. Addressing these will enhance model reliability.

The Random Forest model ( Figure 3 ) showed varying accuracy across Zambia’s ecosystems: strongest in woody savannas (R² = 0.789) where vertical complexity aids prediction, but weak in grasslands (R² = 0.019) and savannas (R² = 0.082) due to limited structure. The combined dataset achieved moderate performance (R² = 0.525), suggesting general applicability despite reduced accuracy at greater heights.

Pixel-level mapping revealed distinct patterns: grasslands (<3 m), woody savannas/Miombo woodlands (up to 30 m). These outputs enable targeted forest management and carbon monitoring by identifying structural changes.

Precipitation showed the strongest positive correlation with height, emphasizing moisture importance. Solar radiation had moderate effects in forests, while temperature relationships were inconsistent.

The Random Forest model demonstrated variable performance across Zambia’s vegetation types ( Figure 4 ). Woody savannas showed exceptional accuracy (R² = 0.789), effectively capturing their complex vertical structure. Combined vegetation types achieved moderate performance (R² = 0.525), suitable for landscape-scale analysis. However, model performance was weak in savannas (R² = 0.082) and grasslands (R² = 0.019), likely due to structural simplicity and data limitations in these ecosystems.

The validated model generated pixel-level height maps across all ecosystems. Results showed distinct patterns: woody savannas exhibited the tallest and most variable canopies, while grasslands remained consistently low. These outputs support targeted forest management and carbon monitoring applications.

Analysis revealed precipitation as the strongest climatic driver of canopy height, particularly in woody savannas. Solar radiation showed moderate positive effects in forested areas, while temperature correlations were generally weak or inconsistent across ecosystems.

While existing forest inventory plots provided useful validation data, their limited spatial coverage may not reflect the full structural variability of Zambia’s ecosystems. A broader field campaign incorporating UAV-LiDAR surveys and strategically located plots is recommended to improve validation accuracy. Moreover, uncertainty propagation throughout the data fusion process remains underexplored. Quantifying the individual contributions of sensors to total model uncertainty would enhance the transparency and robustness of multi-sensor canopy height modelling.

The Random Forest model demonstrated varying performance across Zambia’s vegetation types during training ( Figure 5 ). Woody savannas showed exceptional accuracy (R² = 0.867), with predictions closely matching observations, indicating effective capture of complex vertical structure. Grasslands (R² = 0.769) and savannas (R² = 0.743) also displayed strong training performance, though subsequent validation revealed potential overfitting in these sparser ecosystems. The combined ecosystem model achieved robust performance (R² = 0.76), suitable for landscape-scale applications despite increased variability at greater heights.

When applied spatially, the model generated pixel-level height maps, with vegetation-specific versions improving precision in dominant ecosystems. Climate analysis revealed precipitation and solar radiation as positive drivers of canopy height, particularly in woody savannas, while temperature showed weaker correlations. The woody savanna model emerged as the most reliable, maintaining low RMSE values during validation, unlike grassland models which suffered performance drops due to structural simplicity and heterogeneity.

Three key limitations require attention. First, current validation relies on forest plots that may not represent Zambia’s full ecological diversity—future work should incorporate UAV-LiDAR with stratified field sampling. Second, while accuracy metrics are reported, uncertainty propagation through the multi-sensor fusion process needs quantification using methods like Monte Carlo simulation. Third, the individual contributions of GEDI, Sentinel-1/2, and topographic data to model uncertainty remain unquantified; variance decomposition techniques could clarify each sensor’s impact. Addressing these gaps would enhance the framework’s reliability for operational use across all ecosystems.

The best-performing model ( Figure 6 ) successfully mapped canopy height variability across Zambia, with tallest canopies in Woody Savannas and forest patches. While effective for dense vegetation, its lower accuracy in Grasslands suggests the need for specialized approaches or additional structural metrics in open ecosystems.

Key predictors varied significantly by ecosystem: Woody Savannas relied most on spectral bands (B5, B7, B12), elevation and radar textures, while Grasslands depended primarily on slope and select bands. The combined model consistently prioritized B12, slope and radar data, confirming their cross-ecosystem relevance.

Precipitation and solar radiation showed strong positive correlations with canopy height, particularly in Woody Savannas. Temperature exhibited weaker, inconsistent relationships, suggesting its influence is secondary or ecosystem-specific in shaping vegetation structure.

The current model relies heavily on existing forest inventory plots, which may not fully represent the structural diversity or spatial variability across Zambia. To address this, future work should incorporate more extensive ground validation, including strategically located field plots and UAV-LiDAR surveys. Additionally, the study would benefit from a comprehensive uncertainty analysis, assessing how errors propagate through the data fusion process. Finally, quantifying the individual contributions of each sensor to the model’s uncertainty would enhance transparency and support more robust sensor integration strategies.

The Random Forest model demonstrated strong predictive capability for canopy height estimation in Zambia’s woody savannas training R² = 0.867, validation R² = 0.789, validating its effectiveness in structurally complex ecosystems in Figures 3, 5. This aligns with global studies showing dense vegetation facilitates more accurate remote sensing predictions due to distinct spectral and structural signatures. However, performance declined markedly in grasslands validation R² = 0.019 and savannas (R² = 0.082), reflecting challenges in modeling sparse, low-stature vegetation a limitation noted in similar arid and semi-arid regions. The combined model R² = 0.525 provided reasonable landscape-scale estimates but with heightened variability at taller canopies > 14 m, suggesting ecosystem-specific models may better serve precision applications.

Canopy height distributions in Figure 2 and Figure 6 revealed ecosystem-specific patterns: woody savannas exhibited the tallest (8 - 25 m) and most variable canopies, while grasslands peaked sharply at 4.5 m. These findings corroborate biome-specific growth constraints, with precipitation showing the strongest positive correlation r = 0.56 across all types. Notably, solar radiation’s moderate influence in woody savannas r = 0.43 implies light competition drives vertical stratification in forests, whereas its minimal impact in grasslands underscores herbaceous dominance. Temperature correlations were weak or negative e.g., r = −0.31 in grasslands, suggesting heat stress may suppress growth in moisture-limited systems.

The analysis showed that different ecosystems rely on distinct predictors—woody savannas performed best with combined optical (Sentinel-2 bands), topographic, and radar data, while grasslands depended mainly on terrain features like slope. The consistent importance of the SWIR band (B12) across all ecosystems highlights its usefulness for vegetation moisture detection. However, the model’s uncertainty for sparse vegetation remains unclear due to unquantified sensor-specific contributions. Future work should analyse individual sensor errors to improve accuracy in open ecosystems like grasslands.

While the model demonstrated strong performance in woody ecosystems, several limitations constrain its broader application. The reliance on existing forest inventory plots introduced potential validation bias, particularly for grasslands and savannas where structural diversity was underrepresented. This limitation could be addressed through an integrated validation framework combining UAV-LiDAR surveys with stratified field sampling across all major vegetation types. A second critical gap lies in uncertainty quantification—while the study reported accuracy metrics, it did not track how errors propagate through the multi-sensor data fusion pipeline. Implementing techniques like Monte Carlo simulations would provide essential insights into prediction reliability at pixel scales. Furthermore, the relative contributions of GEDI, Sentinel-1/2, and topographic data to model uncertainty remain unquantified. Future work should employ variance decomposition methods to determine optimal sensor combinations for different ecosystems. These improvements would significantly enhance the model’s operational utility for nationwide canopy height monitoring.

The study’s findings offer both practical applications and research pathways forward. The high-accuracy canopy height maps (R² = 0.789 in woody savannas) provide reliable baselines for carbon accounting and forest management initiatives, particularly in Zambia’s dense woodlands. However, the poor performance in grasslands (R² = 0.019) underscores the need for caution when applying these estimates to open ecosystems—here, the maps may best serve as indicators requiring ground verification. For restoration planning, the identified structural thresholds (e.g., 4.5 m peak in grasslands) can help prioritize intervention areas. From a methodological perspective, the ecosystem-dependent variable importance ( Figure 6 ) suggests future models could benefit from vegetation-specific predictor selection. Two key research priorities emerge: first, incorporating phenological metrics from multi-temporal Sentinel-2 data to account for seasonal vegetation changes; second, testing hybrid modeling approaches that combine Random Forest with deep learning architectures to better capture sparse vegetation dynamics. These advancements would strengthen the model’s utility across Zambia’s diverse biomes while providing a transferable framework for similar savanna-woodland regions globally.