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

Abstract

The models that produce most existing forest height maps minimize mean individual prediction errors quantified by metrics such as Root Mean Square Error (RMSE). However, when predictor data do not explain all height variability in the training sample, this objective function leads to prediction toward the mean, distorting population-level prediction of height range and variability. Several existing forest height maps have used training data from NASA's Global Ecosystem Dynamics Investigation (GEDI) lidar mission, and while the mission's retrievals are subject to measurement error, we used GEDI to: 1) evaluate population-level errors of three prominent global height maps; and 2) produce new maps designed to better represent the full height distribution. Specifically, we introduce the GEDI L4D Imputed Waveform product, which was created using a nearest neighbor algorithm. This model assigned a high-quality waveform, represented by 11 Relative Height metrics and other measurements, to each 30 m pixel in the tropical and temperate domain. Covariates included synthetic Landsat bands derived from time series fit to all clear imagery. While previous maps systematically homogenized tree heights over large regions, L4D did not, and tests verified that L4D preserved appropriate covariance among GEDI-derived forest vertical structure metrics. Validation shows that L4D top height RMSE values were up to approximately 1.1 m greater than other maps. Applications such as predicting timber yield, modeling species distributions, and simulating fire spread may benefit from more demographically accurate tree height maps. Moving forward, attention to population-level accuracy metrics may increase the practical value of remotely sensed maps.