Delineating Standing Deadwood in High-Resolution RGB Drone Imagery

We have observed tree die-offs in a variety of regions in the world. Understanding the diverse causes of tree mortality requires exact information about which trees are dying and where. With the increased user-friendliness of drones and the availability of airborne imagery, high-resolution imagery of forests is becoming widely available. Delineating standing deadwood in such aerial imagery has become a classic segmentation task and several models with varying accuracy have been developed. However, these machine-learning based models are not generic and limited to specific image resolutions, sensor characteristics, geographic regions, and forest ecosystems. The reason for this lack of generality is that previous models have been trained using only datasets representative of specific regions and obtained from a single source. In this study, we obtain a diverse dataset spanning more than a dozen countries across continents and implement a single convolutional neural network (CNN) model that is able to cope with most forest ecosystems, varying image quality, and spatial resolutions.