1. Accurate detection of individual animals is integral to the management of vulnerable wildlife species, but often difficult and costly to achieve for species that occur over wide or inaccessible areas or engage in cryptic behaviours. There is a growing acceptance of the use of drones (also known as unmanned aerial vehicles, UAVs and remotely piloted aircraft systems, RPAS) to detect wildlife, largely because of the capacity for drones to rapidly cover large areas compared to ground survey methods. While drones can aid the capture of large amounts of imagery, detection requires either manual evaluation of the imagery or automated detection using machine learning algorithms.While manual evaluation of drone-acquired imagery is possible and sometimes necessary, the powerful combination of drones with automated detection of wildlife in this imagery is much faster and, in some cases, more accurate than using human observers.Despite the great potential of this emerging approach, most attention to date has been paid to the development of algorithms, and little is known about the constraints around successful detection (P. W. J. Baxter, and G. Hamilton, 2018, Ecosphere, 9, e02194).2. We reviewed studies that were conducted over the last 5 years in which wildlife species were detected automatically in drone-acquired imagery to understand how technological constraints, environmental conditions and ecological traits of target species impact detection with automated methods.3. From this review, we found that automated detection could be achieved for a wider range of species and under a greater variety of environmental conditions than reported in previous reviews of automated and manual detection in droneacquired imagery. A high probability of automated detection could be achieved efficiently using fixed-wing platforms and RGB sensors for species that were large and occurred in open and homogeneous environments with little vegetation or variation in topography while infrared sensors and multirotor platforms were necessary to successfully detect small, elusive species in complex habitats.4. The insight gained in this review could allow conservation managers to use drones and machine learning algorithms more accurately and efficiently to conduct abundance data on vulnerable populations that is critical to their conservation.
Effective management of threatened and invasive species requires regular and reliable population estimates. Drones are increasingly utilised by ecologists for this purpose as they are relatively inexpensive. They enable larger areas to be surveyed than traditional methods for many species, particularly cryptic species such as koalas, with less disturbance. The development of robust and accurate methods for species detection is required to effectively use the large volumes of data generated by this survey method. The enhanced predictive and computational power of deep learning ensembles represents a considerable opportunity to the ecological community. In this study, we investigate the potential of deep learning ensembles built from multiple convolutional neural networks (CNNs) to detect koalas from low-altitude, drone-derived thermal data. The approach uses ensembles of detectors built from combinations of YOLOv5 and models from Detectron2. The ensembles achieved a strong balance between probability of detection and precision when tested on ground-truth data from radio-collared koalas. Our results also showed that greater diversity in ensemble composition can enhance overall performance. We found the main impediment to higher precision was false positives but expect these will continue to reduce as tools for geolocating detections are improved. The ability to construct ensembles of different sizes will allow for improved alignment between the algorithms used and the characteristics of different ecological problems. Ensembles are efficient and accurate and can be scaled to suit different settings, platforms and hardware availability, making them capable of adaption for novel applications.
This thesis advances the way artificial intelligence can be applied to the automatic detection of wildlife from imagery captured with drones. The study represents the first use of deep learning ensembles for this purpose, in which multiple deep learning algorithms of different types and sizes were combined and run simultaneously. The method was able to detect koalas with high precision from drone footage captured in complex eucalypt habitat. A robust method of evaluation was also devised that demonstrated the suitability of the approach, which has strong potential for broader application by providing greater computing power for ecological monitoring.
Light detection and ranging (LiDAR) has been a tool of choice for 3D dense point cloud reconstructions of forest canopy over the past two decades, but advances in computer vision techniques, such as structure from motion (SfM) photogrammetry, have transformed 2D digital aerial imagery into a powerful, inexpensive and highly available alternative. Canopy modelling is complex and affected by a wide range of inputs. While studies have found dense point cloud reconstructions to be accurate, there is no standard approach to comparing outputs or assessing accuracy. Modelling is particularly challenging in native eucalypt forests, where the canopy displays abrupt vertical changes and highly varied relief. This study first investigated whether a remotely sensed LiDAR dense point cloud reconstruction of a native eucalypt forest completely reproduced canopy cover and accurately predicted tree heights. A further comparison was made with a photogrammetric reconstruction based solely on near-infrared (NIR) imagery to gain some insight into the contribution of the NIR spectral band to the 3D SfM reconstruction of native dry eucalypt open forest. The reconstructions did not produce comparable canopy height models and neither reconstruction completely reproduced canopy cover nor accurately predicted tree heights. Nonetheless, the LiDAR product was more representative of the eucalypt canopy than SfM-NIR. The SfM-NIR results were strongly affected by an absence of data in many locations, which was related to low canopy penetration by the passive optical sensor and sub-optimal feature matching in the photogrammetric pre-processing pipeline. To further investigate the contribution of NIR, future studies could combine NIR imagery captured at multiple solar elevations. A variety of photogrammetric pre-processing settings should continue to be explored in an effort to optimise image feature matching.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.