Individual tree detection and classification from RGB satellite imagery with applications to wildfire fuel mapping and exposure assessments

Abstract

Background

Wildfire fuels are commonly mapped via manual interpretation of aerial photos. Alternatively, RGB satellite imagery offers data across large spatial extents. A method of individual tree detection and classification is developed with implications to fuel mapping and community wildfire exposure assessments.

Methods

Convolutional neural networks are trained using a novel generational training process to detect trees in 0.50 m/px RGB imagery collected in Rocky Mountain and Boreal natural regions in Alberta, Canada by Pleiades-1 and WorldView-2 satellites. The workflow classifies detected trees as ‘green-in-winter’/‘brown-in-winter’, a proxy for coniferous/deciduous, respectively.

Key results

A k-fold testing procedure compares algorithm detections to manual tree identification densities reaching an R² of 0.82. The generational training process increased achieved R² by 0.23. To assess classification accuracy, satellite detections are compared to manual annotations of 2 cm/px drone imagery resulting in average F1 scores of 0.85 and 0.82 for coniferous and deciduous trees respectively. The use of model outputs in tree density mapping and community-scale wildfire exposure assessments is demonstrated.

Conclusion & Implications

The proposed workflow automates fine-scale overstorey tree mapping anywhere seasonal (winter and summer) 0.50 m/px RGB satellite imagery exists. Further development could enable the extraction of additional properties to inform a more complete fuel map.