Forest Height Mapping Using Complex-Valued Convolutional Neural Network

Global forest height and biomass mapping is an important issue for ecosystem studies. Polarimetric synthetic aperture radar (SAR) interferometry (PolInSAR) is an attractive technique for extracting forest parameters by inverting the physical scattering model. However, the simplified scattering model restricts the inversion accuracy unless multi-baseline measurements are adopted, and this will lead to increasing cost. In this paper, a complex-valued convolutional neural network (CV-CNN) model using PolInSAR features of single baseline is proposed for forest height mapping with high accuracy. The supervised learning process shows that the obtained CV-CNN model is accurate enough to describe the complicated forest scattering process. Both simulations and experiments on airborne E-SAR data set demonstrate that the proposed CV-CNN model-based method can greatly improve the forest height inversion accuracy. Experimental results show that the coefficient of determination (r2) increases from 0.83 to 0.92, while root-mean-square error (RMSE) decreases from 3.74m to 2.58m. This promising approach makes it possible to map forest heights more accurately from the single-baseline PolInSAR observations, which will further promote the wide use of PolInSAR technique.