High-Accuracy Canopy Height Mapping by Fusing Satellite LiDAR and Imagery through Deep Learning

Abstract

Forests play an important role in mitigating and adapting to global warming, and forest quality may be more important than the amount of forest. Therefore, there is global demand to create a method to monitor canopy height and forest vertical structure. Satellite Light Detection and Ranging (LiDAR) with satellite imagery can produce a map that describes the forest quality. High-accuracy mapping is expected to be achievable through deep learning. However, over-training may cause overfitting. In this study, wall-to-wall forest canopy height maps were created by fusing Global Ecosystem Dynamics Investigation Level 2A (GEDI L2A) data on canopy height (rh_095) and PlanetScope satellite images, to create data for use by a Convolutional Neural Network (CNN)-based deep learner. An early stopping method was introduced in this study to stop training before the learning process starts to produce large canopy height errors. In order to avoid inappropriate accuracy assessment due to bias in the validation sites, forest canopy height maps were validated for the entire area. The true value was taken as the result of satellite LiDAR waveform simulation using Airborne Laser Scanning (ALS) point cloud data. The tree height estimation error in the present study area in the Izu Peninsula had a Root Mean Square Error (RMSE) of 4.94 m; that is, the estimation error was significantly reduced by more than 1.7 m. Initially, 25% of the training data used in this study contained large height errors, however, our early stopping method successfully eliminated the pre-screening process of the error data . Furthermore, external reference data were no longer required during data screening. Our proposed method also automatically determines when to stop learning. This method could save a significant amount of the time and computational resources required to produce high-accuracy global tree height maps.