Beyond RMSE: The GEDI Imputed Waveform product minimizes artificial homogeneity evident in existing global forest height maps

The models that produce most existing forest height maps minimize the mean individual prediction errors that are quantified by metrics such as Root Mean Square Error (RMSE). However, to whatever degree the predictor data do not explain all of the height variability in the training sample, this objective function necessarily leads to prediction toward the mean at the expense of distorting population-level prediction of height range and variability. Several existing forest height maps have used training data from NASA’s GEDI lidar mission, and while the Mission’s retrievals are subject to measurement error, we too used GEDI to both: 1) evaluate population-level errors of three prominent global height maps; and 2) produce a new set of maps designed to more accurately represent the full distribution of heights. Specifically, we created the GEDI L4D Imputed Waveform product, which assigns a high-quality waveform and all constituent measurements to each 30m pixel in the tropical and temperate latitudes that GEDI measures. While the previous maps systematically homogenized tree heights over large regions, L4D did not, and tests verified that L4D preserved appropriate covariance among the many variables measured by GEDI. L4D top height RMSE values were on the order of only 1m greater than other maps. Applications as diverse as predicting timber yield, modeling species distribution, and simulating fire spread rates may benefit from more demographically accurate tree height maps. As modeling methods evolve, attention to population-level accuracy metrics may increase the practical value of remotely sensed maps.