Estimation of Broadleaf Tree Canopy Height of Wolong Nature Reserve Based on InSAR and Machine Learning Methods

Abstract

Tree height is an important parameter for calculating forest carbon sink and assessing forest carbon cycle. In order to obtain forest tree height over a large area both efficiently and at a low cost, this study proposed an Interferometric Synthetic Aperture Radar (InSAR) combined with a machine learning method to estimate the tree canopy height. The forest height in the study area was obtained using Unmanned Aerial Vehicle (UAV) photogrammetry, which was considered to be the true canopy height. Two machine learning methods (Random Forest, Multi-layer perceptron) were used to establish the relationship between phase center height calculated by InSAR DEM differential interference method and coherent amplitude method with true canopy height. The topographic factor, backward scattering coefficient and coherence coefficient were introduced into the relationship model. It was found that the accuracy of tree height estimation using random forest and two InSAR methods can reach 0.95 and 0.94. The root-mean-square error was 1.76 m, 1.86 m, respectively. The accuracy of tree height estimation using multi-layer perceptron and two InSAR methods was 0.25 and 0.2. The root-mean-square error was 3.96 m and 4.13 m. The results indicated that the combination of InSAR and machine learning can estimate canopy height efficiently and at a low cost. Moreover, the integrated learning algorithm random forest demonstrated better stability and higher accuracy than the single learning algorithm multi-layer perceptron.