Using deep learning to detect an indicator arid shrub in ultra-high-resolution UAV imagery

•Deep-learning-based object detection is feasible for arid conservation monitoring.•Accuracies of 75–85 % for pearl bluebush detection.•No loss of accuracy when using lower GSD imagery (3 cm vs 0.8 cm).•Allows for 16 × greater imaging area in same time (3 cm vs 0.8 cm).•No gain in accuracy by using a ResNet CNN with more than 50 layers. Effective monitoring of arid and semi-arid rangelands around the world is essential to understand and combat degradation caused by anthropogenic use and facilitate effective management practices. Remote sensing technologies provide ideal approaches for enhancing traditional on-ground monitoring. However, while broad-scale monitoring of vegetation in rangelands using satellites has been widely adopted, there has been far less uptake of remote sensing for measuring fine-scale indicators of ecosystem condition. This study demonstrates the feasibility of using ultra-high-resolution UAV (Uncrewed Aerial Vehicle) imagery and deep-learning-based object detection models to provide plant recognition and survey information relevant for operational monitoring programmes in arid and semi-arid ecosystems. Seven different object detectors using varying convolutional neural network (CNN) architectures are tested at three image resolutions to detect a widespread, dominant arid shrub species (pearl bluebush, Maireana sedifolia) that serves as a key indicator of overall site condition in southern Australian rangelands. To maximise the strength of statistical analysis, each method is trained on six different training datasets (each using 2,000 to 3,000 training samples) at six widely dispersed sites. This results in 90 trained models, each validated at two sites. To test model generalisability, training and validation data was always sourced from separate vegetation monitoring sites. The influence of variability between sites on detection accuracy is also considered. The best performing models achieved F1 scores (overall accuracy) of around 75% for pearl bluebush detection, a level of accuracy that provides useful monitoring information to land managers. Information extracted from UAV imagery using this approach relates directly to indicators of ecological condition measured in ground-based monitoring; including dominant plant species count, location and density. Continued development and eventual implementation of this method would provide objective conservation-relevant information at broad scales in a far reduced time and at a lower cost t