A Machine-Learning Approach to PolInSAR and LiDAR Data Fusion for Improved Tropical Forest Canopy Height Estimation Using NASA AfriSAR Campaign Data

This paper investigates the benefits of integrating multi-baseline polarimetric interferometric SAR (PolInSAR) data with LiDAR measurements using a machine learning approach in order to obtain improved forest canopy height estimates. Multiple interferometric baselines are required to ensure consistent height retrieval performance across a broad range of tree heights. Previous studies have proposed multi-baseline merging strategies using metrics extracted from PolInSAR measurements. Here, we introduce the multi-baseline merging using a Support Vector Machine trained by sparse LiDAR samples. The novelty of this method lies in the new way of combining the two datasets. Its advantage is that it does not require a complete LiDAR coverage, but only sparse LiDAR samples distributed over the PolInSAR image. LiDAR samples are not used to obtain the best height among a set of height stacks, but rather to train the retrieval algorithm in selecting the best height using the variables derived through PolInSAR processing. This enables a more accurate height estimation for a wider scene covered by the SAR with only partial LiDAR coverage. We test our approach on NASA AfriSAR data acquired over tropical forests by the L-band UAVSAR and the LVIS LiDAR instruments. The estimated height from this approach has a higher accuracy (r 2=0.81, RMSE = 7.1 m) than previously introduced multi-baselines merging approach (r 2=0.67, RMSE = 9.2 m). This method is beneficial to future spaceborne missions such as GEDI and BIOMASS, which will provide a wealth of near-contemporaneous LiDAR samples and PolInSAR measurements for mapping forest structure at global scale.