Published November 12, 2026 | Version v1
Dataset Restricted

2020 Forest structure and biomass at 30-m resolution for California using GEDI and Google Alpha Earth Embeddings

  • 1. EDMO icon University of Maryland
  • 2. EDMO icon National Aeronautics and Space Administration, Jet Propulsion Laboratory

Description

Annual AlphaEarth Foundations (AEF) satellite embeddings were obtained from the Google Earth Engine collection GOOGLE/SATELLITE_EMBEDDING/V1/ANNUAL for 2018. First, training data calibration model scoped for the entire state of California was derived using same crossovers between GEDI footprint-level metrics on canopy height (Relative Height 98 in L2A product), canopy cover (L2B) and aboveground biomass density predictions available in the GEDI L4A product moduled by plant functional type. We used AEF scenes and GEDI waveform temporally co-registered in year 2020. The year choice is explained by higher altitude conditions at the ISS than normal after 2020 which resulted in less separation between GEDI ground tracks and slowed down the gain in coverage. In addition, GEDI only partially collected data in 2019 so the choice of 2020 was the best sounded option within the 2017-2025 period available for AEF. In terms of spatial resolution, the maps aimed to inform LANDFIRE whose data layers are scoped at 30 m resolution. AEF images were upscaled to 30-m resolution and mosaicked into a 64-band predictor image – all native AEF bands were retained to generate the training data.

Pairs between AEF and GEDI footprint locations were driven by pixel-to-waveform co-registration, high-quality filtering and balanced sampling across all counties in California - a large state rich in diverse forests, climate and topographical conditions. Quality filters in the three GEDI products were activated when retrieving high-fidelity footprints complemented by filters of signal degradation (degrade_flag = 0, best quality) and topography – GEDI measurements in slopes above 30 degrees were not considered as these can induce noise particularly over short-stature and sparse forests. The extreme filtering reduced the actionable GEDI archive but this was not a limitation to ensure that at least 5,000 observations passed the filter on each of California’s 58 counties. The search operation was repeated to a maximum of 10 times on each state to increase the number and spatial representativeness of the sample. In this process, we aimed to generate a spatially-representative dataset with balanced observations across the domain of the GEDI variables. Candidate footprints were classified into 5-m height, 10% cover and 50 Mg ha-1 classes before retrieving observations from each class. To ensure representatives and to mitigate the underrepresentation of tall-tree, high-biomass forests during the search – we looked for GEDI footprints across shrublands, grasslands and forests – we ensured that at least 2,500 GEDI observations showed canopy height > 50 m, canopy cover > 95% and biomass > 900 Mg ha-1. For other classes in each GEDI metric, the sampling was progressively adjusted to ensure balanced representativeness – by domain and by county – to conform a dataset of 40,000 observations. Because we strive for science reproducibility and transparent outreach, all steps were conducted within GEE. Preliminary tests revealed that 50,000 was the maximum number of observations allowed to run the calibration model between GEDI and AEF – and that explains the strategy used to generate the calibration database.

Each GEDI variable response was modeled using the Smile gradient tree boosting implementation available through in GEE, configured for regression with a least-squares loss. The single fitted model used 150 trees, a shrinkage parameter of 0.05, a sampling rate of 0.7, a maximum of 32 terminal/decision-tree nodes as specified by maxNodes, and random seed 100. The trained model was applied pixelwise to the reduced embedding image to generate a continuous canopy height, canopy cover and aboveground biomass density prediction surface.

 We used 70% partition for model training development and 30% for holdout testing. To characterize sensitivity to the training subset and random initialization, 50 additional gradient tree boosting regressors were fitted. For each ensemble member, approximately 80% of the original training partition was selected to fit a model using the same hyperparameters as the single model. Each model generated a prediction image, and the 50 images were aggregated pixelwise to obtain the ensemble mean and standard deviation. The mean was interpreted as the ensemble estimate, whereas the standard deviation quantified between-model prediction dispersion. A coefficient of variation was calculated as the standard deviation divided by the absolute ensemble mean, with a minimum denominator of 0.0001 to avoid division by zero. The 5th and 95th percentiles of ensemble predictions were also derived to summarize their empirical spread. Predictive performance for each GEDI metric of interest was evaluated on the 30% holdout subset, using predictions from the single gradient tree boosting model rather than from the 50-member ensemble.

Mathematical definitions for accuracy statistics

·       For pixel x, the ensemble prediction is μ(x) = (1/M) Σₘ ŷₘ(x), with M = 50.

·       Between-model standard deviation is s(x) = √[(1/(M − 1)) Σₘ (ŷₘ(x) − μ(x))²], following the standard-deviation reducer convention.

·       Coefficient of variation is CV(x) = s(x) / max(|μ(x)|, 0.0001).

·       RMSE = √[(1/n) Σᵢ (ŷᵢ − yᵢ)²]; MAE = (1/n) Σᵢ |ŷᵢ − yᵢ|; R² = 1 − [Σᵢ (ŷᵢ − yᵢ)² / Σᵢ (yᵢ − ȳ)²].

Accuracies

The RMSE, MAE and R² for canopy height were 9.92 m, 6.6 m and 0.63 respectively.

The RMSE, MAE and R² for canopy height were 0.23, 0.18 and 0.48 respectively.

The RMSE, MAE and R² for canopy height were 139.6 Mg ha-1, 99.7 Mg ha-1 and 0.57 respectively.

 

Files

Restricted

The record is publicly accessible, but files are restricted. Log in to check if you have access.

Additional details

Dates

Available
2020-01-01/2020-12-31