Synthesizing Disparate LiDAR and Satellite Datasets through Deep Learning to Generate Wall-to-Wall Regional Forest Inventories

Light detection and ranging (LiDAR) has become a commonly-used tool for generating remotely-sensed forest inventories. However, LiDAR-derived forest inventories have remained uncommon at a regional scale due to varying parameters between LiDAR datasets, such as pulse density. Here we develop a regional model using a three-dimensional convolutional neural network (CNN), a form of deep learning capable of scanning a LiDAR point cloud as well as coincident satellite data, identifying features useful for predicting forest attributes, and then making a series of predictions. We compare this to the standard modeling approach for making forest predictions from LiDAR data, and find that the CNN outperformed the standard approach by a large margin in many cases. We then apply our model to publicly available data over New England, generating maps of fourteen forest attributes at a 10 m resolution over 85 % of the region. Our estimates of attributes that quantified tree size were most successful. In assessing aboveground biomass for example, we achieved a root mean square error of 36 Mg/ha (44 %). Our county-level mapped estimates of biomass were in good agreement with federal estimates. Estimates of attributes quantifying stem density and percent conifer were moderately successful, with a tendency to underestimate of extreme values and banding in low density LiDAR acquisitions. Estimate of attributes quantifying detailed species groupings were less successful. Ultimately we believe these maps will be useful to forest managers, wildlife ecologists, and climate modelers in the region.