Silvilaser 2019 - Poster Presentations »
Individual crown detection and height measurement of western juniper with deep convolutional and generative adversarial neural networks
In sparse and irregular forest, like those found within the juniper savannah in Oregon, the effectiveness of standard forest inventories is reduced. Instead, for large area evaluation, the usage of remotely sensed data is recommended to estimate the attributes of interest. Western juniper (Juniperus occidentalis Hook.) is a native invasive species in Oregon, which increased its area five-fold since 1930. Juniper landscape restorations efforts are challenged by its sparse spatial distribution (trees per hectare). The objective of this study is to help the restoration process by estimating the location and size of western junipers, with an application to Wheeler County in eastern Oregon. We focused on individual tree crown detection at the landscape level using two approaches, one based on Faster RCNN (region-based convolutional neural network) and one on a watershed detection algorithm augmented by a canopy height model generated from a generative adversarial network (GAN). Both approaches identified more than 70% of the trees, with Fast RCNN outperforming the combination GAN-watershed algorithm, with an omission error 7% and commission error 26%. The inclusion of the canopy height model improved tree crown identification approximately 15%, revealing the benefits of using GAN. Besides providing individual tree locations, the Fast RCNN reduced tree identification errors by almost half compared with the existing results. Our findings suggest that estimation of 3D characteristics of a forest with GAN has significant potential in forest inventory and can improve existing methods of estimation of canopy metrics. Furthermore, our results support the usage of Faster RCNN in detecting tree crowns, and, possibly, other forest relevant objects, such as coarse woody debris, from remotely sensed data.