Mapping Vegetation Structure from UAVSAR Tomography Using 3-D Convolutional Neural Networks

The NASA/JPL Uninhabited Aerial Vehicle Synthetic Aperture Radar (UAVSAR) instrument has performed tomographic SAR experiments over a number of study areas, including Rabi Forest in Gabon in 2016 and Sierra National Forest in California, USA in 2021. Tomographic SAR, or TomoSAR, is a technique enabling 3-D radar imaging with diverse applications including mapping of vegetation structure. Convolutional neural networks (CNNs) have shown widespread potential for many image processing and computer vision tasks such as image segmentation, classification, and object recognition. By using 3-D CNNs rather than 2-D CNNs, the filters can be applied to all three dimensions of a forest volume imaged by TomoSAR. We have trained 3-D CNN-based deep learning models to estimate canopy height and canopy cover from fully polarimetric UAVSAR TomoSAR images using lidar data as training and validation. When applied to canopy height estimation in the Rabi Forest study area, a trained network had root mean square error (RMSE) of 2.41 m (7\%) for test data held out from training. For canopy cover estimation in the Sierra National Forest study area, a trained network yielded RMSE of 10\% for test data held out from training. Further work can be done to optimize the network architecture, improve the output spatial resolution, and to check if these methods can be applied to other study areas or to other vegetation structure parameters such as above-ground biomass. The results show the strong potential of 3-D CNNs for mapping wall-to-wall vegetation structure from tomographic SAR imagery using lidar training data.