CANCELLED - Forest monitoring via segmentation and classification of tree species in a Bottomland Hardwood Forest using a U-Net Convolutional Neural Network

Background/Question/Methods

Consumer Unmanned Aerial Systems (UAS) hold great promise in the remote sensing of vegetation at both small and large scales (< 10 hectares). In this study, we use an off-the-shelf UAS to accurately segment and classify tree canopies within a mature Bottomland Hardwood Forest (BLHF) in Northeast Louisiana. We use nine classes, representing all dominant taxa (Quercus, Fraxinus, Gleditsia, Carya, Ulmus, Salix, and Celtis) and relevant physiognomic indicators (canopy gaps and understory). Classes were manually digitized on an orthomosaic photo and ground-truthed in the field on a mobile GIS application. A five-band image product was created using a Canopy-Height Model and Grey-Scale Morphological Texture index in addition to the RGB bands acquired using UAS. Pixel-wise image segmentation was conducted using a U-Net Convolutional Neural Network to assign class values in a manner that accounts for spatial relationships between neighboring features. Our model attempts to classify a comparatively complex class structure and study system while maintaining well-defined canopy representations. Conducting vegetation surveys in a BLHF are often treacherous due to the hydrologic influences. UAS technology allows us to confidently demarcate individual trees and assess compositional changes of vegetation over time in the face of stochastic hydrology and anthropogenic influence.

Results/Conclusions

A total of 1,423 individual tree canopies were delineated for training, validation, and test datasets across 6.6 hectares in Russel Sage Wildlife Management Area. The dominant genus in the forest is Quercus sp., representing 72.3% of individuals and 80.5% of canopy area. Precision and recall are led in accuracy by the two physiognomic classes and Gleditsia sp. with a maximum value of 97.0%. Quercus sp., Fraxinus sp., and Carya sp. all have overall accuracies above 60%. The limited sample set of Salix nigra (3 individuals) and Celtis laevigata (14 individuals) results in a poor accuracy of 2% and 37% respectively. Visually, individual canopies are homogenous and show strong discrimination amongst classes. Preliminary results show success in delimiting species of trees in a BLHF through current means, however differential band combinations, texture indices, and advanced network architectures are being explored to increase model accuracy and discriminatory power for species identification. Understanding canopy and species dynamics temporally will allow land managers to respond to emerging threats such as the Emerald Ash Borer, Oak Decline, or extreme weather events when evaluating stands and preparing justified responsive actions at scale.