GEDI: Global Ecosystem Dynamics Investigation

• GEDI is a NASA mission employing full-waveform LiDAR to capture high-resolution vertical forest structure data, crucial for monitoring terrestrial carbon cycles.
• It integrates advanced machine learning, notably 1D CNNs with residual blocks, to accurately estimate canopy height and quantify both measurement and model uncertainty.
• GEDI’s refined canopy height mapping reduces RMSE from ~4.4 m to 2.7 m, enabling robust calibration of global ecosystem models and improved biomass quantification.

The Global Ecosystem Dynamics Investigation (GEDI) is a NASA mission that employs spaceborne full-waveform LiDAR to acquire high-resolution vertical forest structure data for advancing terrestrial carbon cycle science, ecosystem modeling, and climate research. Positioned on the International Space Station since 2018, GEDI’s explicit focus is on the retrieval of vegetation height, canopy structure, and aboveground biomass—parameters essential for global carbon budget quantification and forest management. GEDI serves as a reference standard for calibrating and validating global ecosystem models and enhances the precision of large-scale remote sensing applications through robust physical measurements.

1. Instrumentation and Measurement Principles

GEDI’s instrumentation is based on a full-waveform LiDAR system that operates with a fixed waveform length (1420 vertical bins; ≈15 cm vertical bin spacing; ~25 m footprint diameter), delivering high-fidelity vertical returns across ecologically diverse terrain. The core measurement is a footprint-level waveform that records the amplitude of reflected laser energy as a function of vertical distance, enabling detection of both upper canopy and ground returns under a wide range of environmental and observational conditions.

The waveform is preprocessed (including background noise subtraction and energy normalization), then analyzed to extract biophysically meaningful metrics such as relative height (RH) percentiles—especially RH98, the 98th percentile height, widely used as a proxy for maximum canopy height in ecological and biometrical applications (Lang et al., 2021). GEDI’s spatial sampling strategy covers approximately 4% of the nominal land surface between 51.6°N/S with a sparse, non-repeating grid, making generalized or wall-to-wall mapping a non-trivial downstream problem.

2. Machine Learning Algorithms for Canopy Height Estimation

Interpreting the voluminous GEDI waveform data over heterogeneous landscapes requires automated, scalable approaches. Deep learning—especially supervised convolutional neural networks (CNNs)—has been demonstrated as an effective solution for regression of biophysical variables directly from the raw waveform data (Lang et al., 2021).

• The most prominent architecture is a 1D CNN with eight residual blocks (each block: two consecutive 1D convolutions, batch normalization, ReLU activations, skip connections), inspired by ResNet. Max-pooling layers are inserted between blocks to improve the receptive field and contextualize vertical energy distributions.
• The trained model ingests noise-subtracted, fixed-length, energy-normalized waveforms and produces as output both a mean canopy height estimate (the predicted RH98) and, crucially, an associated predictive variance.
• This design allows the network to directly learn physically interpretable features (canopy peaks, ground returns, vertical attenuation patterns) that are robust to region-specific or noise-specific waveform artifacts. Training data, sourced from 68,483 globally distributed footprints, are co-located with airborne laser scanning (ALS) reference data to provide reliable ground truth for supervised learning. Generalization across biomes and continents is rigorously quantified through geographic holdout cross-validation.

3. Probabilistic Prediction and Uncertainty Quantification

The GEDI deep learning framework is probabilistic, capturing both aleatoric uncertainty (measurement and environmental noise) and epistemic uncertainty (model uncertainty arising from training data limitations or distributional shift). This is achieved via:

• Each CNN is trained to output both a mean (𝜇̂) and variance (𝜎̂²) prediction for the latent Gaussian conditional distribution, and optimized via the negative log-likelihood (NLL) loss:

$$\mathcal{L}_{NLL} = \frac{1}{N} \sum_{i=1}^N \left[ \frac{ \big( \hat{\mu}(x_i) - y_i \big)^2 }{2\,\hat{\sigma}^2(x_i)} + \frac{1}{2} \log \hat{\sigma}^2(x_i) \right]$$

• An ensemble of ten independently-initialized CNNs provides sample-based estimates of epistemic uncertainty. For a test waveform, the final mean and variance are computed as:
  • Mean: $\hat{\mu} = \frac{1}{M} \sum_{m=1}^M \hat{\mu}_m$
  • Total predictive variance: $$\operatorname{Var}(\hat{y}) = \frac{1}{M}\sum_{m=1}^M \hat{\mu}_m^2 - \left(\frac{1}{M}\sum_{m=1}^M \hat{\mu}_m\right)^2 + \frac{1}{M}\sum_{m=1}^M \hat{\sigma}_m^2$$

This systematic uncertainty quantification enables not only the filtering of low-confidence predictions but also the propagation of error estimates for downstream carbon and ecological modeling.

4. Feature Extraction, Generalization, and Model Robustness

The convolutional filters and residual structure of the network ensure that the learned representations capture both fine-scale and broad-scale waveform features. This property is critical for generalization:

• Local details—such as sharp peaks corresponding to canopy tops or abrupt ground returns—are captured by early convolutional layers.
• Max-pooling and skip connections allow aggregation of extended vertical context, crucial for discriminating overlapping canopy/ground waveforms in sloped or complex vegetation.
• Training on a globally diverse, ALS-validated footprint set and implementing geographic cross-validation (whereby the model is withheld from a region during training and tested on that region) confirm that the network does not overfit to regional signature artifacts or instrument-specific quirks but instead learns genuinely transferable biophysical signals.

5. Model Performance and Calibration

The effectiveness of the approach is quantified using standard metrics:

Metric	Formula	Interpretation/Result
RMSE	$$\sqrt{ \frac{1}{N}\sum_{i=1}^N (\hat{y}_i - y_i)^2 }$$	As low as 2.7 m (filtered high-confidence)
ME	$\frac{1}{N} \sum_{i=1}^N (\hat{y}_i - y_i)$	–0.3 m overall (–0.1 m for filtered subset)
Calibration	Compare predicted $\hat{\sigma}$ with empirical RMSE in bins of uncertainty	Close match, indicating realistic estimates
MAE, MAPE	Standard absolute/percentage errors for additional characterization of model deviation	Noted in results but secondary to RMSE

The substantial reduction of RMSE compared to earlier GEDI algorithms (from ∼4.4 m to 2.7 m) underscores the impact of deep ensemble learning paired with uncertainty modeling. Moreover, the near-zero mean error validates the absence of systematic bias after careful filtering and uncertainty calibration.

6. Implications for Carbon Cycle and Ecosystem Monitoring

Accurate mapping of canopy top height has direct implications for quantifying aboveground biomass, regional and global carbon stock assessment, and carbon flux modeling. The infusion of predictive uncertainty enables the filtration and weighting of GEDI-derived estimates in downstream applications:

• Forest carbon pool estimation is improved—both in mean and error variance terms—by deploying only highly confident predictions, supporting more robust climate model constraints and regional carbon sink/source quantification.
• Biomass mapping pipelines and ecosystem models can directly propagate the well-calibrated uncertainty to derive conservative or risk-informed metrics for management and policy.
• By substantially reducing bias and error in canopy height estimates, the methodology enables a reduction in the uncertainties that dominate global carbon budget calculations—a central mission objective for GEDI.

This evidence-based deep learning approach illustrates the complementarity of modern probabilistic learning methods with physical remote sensing for planetary-scale ecosystem monitoring, advancing both algorithmic and practical standards for the interpretation of spaceborne LiDAR data in carbon cycle science.