Tropical Forest Top Height by GEDI: From Sparse Coverage to Continuous Data

Abstract

Estimating consistent large-scale tropical forest height using remote sensing is essential for understanding forest-related carbon cycles. The Global Ecosystem Dynamics Investigation (GEDI) light detection and ranging (LiDAR) instrument employed on the International Space Station has collected unique vegetation structure data since April 2019. Our study shows the potential value of using remote-sensing (RS) data (i.e., optical Sentinel-2, radar Sentinel-1, and radar PALSAR-2) to extrapolate GEDI footprint-level forest canopy height model (CHM) measurements. We show that selected RS features can estimate vegetation heights with high precision by analyzing RS data, spaceborne GEDI LiDAR, and airborne LiDAR at four tropical forest sites in South America and Africa. We found that the GEDI relative height (RH) metric is the best at 98% (RH98), filtered by full-power shots with a sensitivity greater than 98%. We found that the optical Sentinel-2 indices are dominant with respect to radar from 77 possible features. We proposed the nine essential optical Sentinel-2 and the radar cross-polarization HV PALSAR-2 features in CHM estimation. Using only ten optimal indices for the regression problems can avoid unimportant features and reduce the computational effort. The predicted CHM was compared to the available airborne LiDAR data, resulting in an error of around 5 m. Finally, we tested cross-validation error values between South America and Africa, including around 40% from validation data in training to obtain a similar performance. We recommend that GEDI data be extracted from all continents to maintain consistent performance on a global scale. Combining GEDI and RS data is a promising method to advance our capability in mapping CHM values.

Evaluating GEDI data fusions for continuous characterizations of forest wildlife habitat

“…Similar to our results, Ngo et al (2023) also found optical data the most important predictors of RH98 even when radar was included, with a validation R 2 of 0.62 and RMSE of 5 m when compared to an ALS canopy height. All previous studies of scaled-up GEDI RH98 predictions reviewed here had similar biases to those from our study, with under predictions at taller heights and over predictions at lower RH98 values (Healey et al, 2020;Sothe et al, 2022;Ngo et al, 2023).…”

“…Our results showed promise for applying GEDI-fusion models further back in time than our study period starting in 2016, where we found comparable map validation results for hindcasted years versus years of model creation for all structure metrics. We also found comparable performance for models based on Landsat, topographic, and bioclimatic predictors compared to the model that also incorporated Sentinel-1 and disturbance metrics, which supports the findings of Ngo et al (2023) for the importance of optical data for scaling up GEDI information to continuous extents. The reliance on Landsat data for driving our GEDI-fusion models provides the opportunity to hindcast models further back across the Landsat archive prior to the availability of Sentinel-1 data.…”

“…At this 3 km local calibration scale, they were able to achieve an RMSE of 7.08 m for RH98 predictions using only Landsat predictors across globalextents (Healey et al, 2020), which shows promise for the potential of expanding GEDI structure measures to larger extents than included in our regional study. Within sample tropical forest sites in South America and Africa, Ngo et al (2023) found the greatest modeling success with RH98 among the possible upper canopy height GEDI metrics, supporting our inclusion of this variable. Similar to our results, Ngo et al (2023) also found optical data the most important predictors of RH98 even when radar was included, with a validation R 2 of 0.62 and RMSE of 5 m when compared to an ALS canopy height.…”

“…The consistency between footprint-and maplevel assessments further supports the promise of data fusion approaches for scaling up GEDI structure information to continuous extents, even with geolocation errors from both GEDI footprints (e.g., up to 10 m error within version 2 GEDI data) and satellite predictors (e.g., potential of 15 m error for Landsat pixels). Among the studies focused on scaling up GEDI structure metrics through data fusion approaches, the majority have largely focused upon GEDI canopy heights (Healey et al, 2020;Sothe et al, 2022;Ngo et al, 2023). Similar to these previous studies, we found high accuracies within our maps scaling up RH98 samples to continuous extents through optical and radar-based data fusions with an R 2 of 0.673 and RMSE of 6.996 m when compared against simulated ALS validation samples, and an R 2 of 0.757 and RMSE of 5.445 m when assessed using withheld GEDI footprint testing data.…”

Women in remote sensing: 2022

“…Among the studies focused on scaling up GEDI structure metrics through data fusion approaches, the majority have largely focused upon GEDI canopy heights (Healey et al, 2020;Sothe et al, 2022;Ngo et al, 2023). Similar to these previous studies, we found high accuracies within our maps scaling up RH98 samples to continuous extents through optical and radar-based data fusions with an R 2 of 0.673 and RMSE of 6.996 m when compared against simulated ALS validation samples, and an R 2 of 0.757 and RMSE of 5.445 m when assessed using withheld GEDI footprint testing data.…”