2 min readfrom Frontiers in Marine Science | New and Recent Articles

Using semi-supervised learning to detect beluga whales from aerial image sequences

Our take

Precise beluga whale population monitoring is essential for Arctic conservation, yet manual annotation of aerial imagery presents significant logistical and cost barriers. This study investigates a solution: semi-supervised learning. We systematically evaluated SEMI-DETR, a novel detection transformer, against supervised methods, demonstrating a 20% mean Average Precision improvement with just 1% of labeled data. Notably, calf detection—critical for assessing reproductive health—benefited most. These findings establish an empirical benchmark and offer practical guidance for resource-constrained conservation programs, mirroring approaches explored in our related work on plankton monitoring.
Using semi-supervised learning to detect beluga whales from aerial image sequences

The challenge of accurately and efficiently monitoring wildlife populations, particularly in remote and challenging environments like the Arctic, is a persistent hurdle for conservation efforts. Precise data on population size, age structure, and reproductive health are vital for informed management decisions, yet traditional methods relying on manual annotation of imagery are both time-consuming and expensive. This new research, employing semi-supervised learning to detect beluga whale calves from aerial drone imagery, offers a significant step forward in addressing this bottleneck. The work builds upon advancements in automated image analysis, as demonstrated by similar efforts in plankton classification [Plankton imager 10 monitoring in the southern North Sea: an open workflow for classification, morphometry and DwC-A publication], and highlights the increasing potential of leveraging machine learning to enhance our understanding of marine ecosystems. Furthermore, the difficulties of operating real-time object detection systems in degraded underwater environments—a challenge explored in [Reliability-aware query restoration for embedded real-time object detection in degraded underwater vision systems]—underscores the ingenuity required in developing robust automated monitoring solutions.

The core innovation here lies in the application of SEMI-DETR, a semi-supervised detection transformer, which dramatically reduces the need for manually labeled data. The study’s finding that comparable detection performance can be achieved with a 5- to 10-fold reduction in annotation requirements is particularly noteworthy. Calf detection, notoriously difficult due to their smaller size and variable appearance, showed a substantial 24.3% relative improvement with semi-supervised learning at the 1% labeling level. This is a crucial advancement, as the health and survival rates of beluga calves are direct indicators of population viability. The research provides an empirical benchmark for semi-supervised marine mammal detection, offering valuable guidance for resource allocation in conservation programs. It's a practical demonstration of how intelligent algorithms can augment, not replace, the expertise of human observers, allowing them to focus on more complex analytical tasks. The work’s design, utilizing a bespoke dataset of 7,655 high-resolution aerial images, demonstrates a commitment to rigorous evaluation and provides a foundation for future studies in similar contexts.

Beyond the immediate implications for beluga whale conservation, this research has broader relevance for wildlife monitoring across diverse species and habitats. The principle of minimizing annotation burden while maintaining detection accuracy is applicable to a wide range of ecological studies, from tracking endangered species to assessing the impact of environmental changes on biodiversity. The success of SEMI-DETR highlights the growing importance of exploring semi-supervised and unsupervised learning approaches in conservation science. The challenges faced in monitoring dugong populations in Shark Bay, as detailed in [Wuthuga (dugong) population status in Gathaagudu (Shark Bay) sea country], demonstrate the broader need for innovative and efficient monitoring techniques, and this study provides a powerful tool for addressing those needs. The integration of real-time data streams and sophisticated algorithms will be critical for proactive conservation efforts.

Looking ahead, the development of increasingly sophisticated semi-supervised learning models will likely revolutionize wildlife monitoring. The ability to train robust detection systems with limited labeled data opens up new possibilities for monitoring remote and understudied populations. A key question will be how to effectively combine these automated systems with traditional ecological knowledge, ensuring that conservation strategies are both scientifically sound and culturally sensitive. Furthermore, the integration of longitudinal data streams—measuring climate indicators and other environmental factors—will provide a more holistic understanding of the drivers influencing population dynamics and allow for adaptive management strategies. The convergence of advanced data analytics and ecological expertise promises a new era of informed and effective conservation.

IntroductionPrecise population monitoring of beluga whales (Delphinapterus leucas) is crucial for Arctic conservation management; yet, the manual annotation of aerial drone imagery is costly and logistically challenging. Calf detection is particularly important because calf presence and survival rates are indicators of reproductive health, recruitment success, and long-term population viability. However, calves are challenging detection targets due to their smaller size, lower frequency, and greater variability in appearance compared with adults. Although deep learning-based object detection offers considerable potential for automated wildlife monitoring, supervised approaches typically require large quantities of manually annotated data, limiting their application in conservation programs with restricted annotation resources.MethodsThis study systematically evaluated semi-supervised marine mammal detection from aerial imagery by comparing SEMI-DETR, a semi-supervised detection transformer, with four variants of YOLO11 across seven annotation budgets ranging from 1% to 50% of labeled data. Experiments were conducted using a bespoke dataset of 7,655 high-resolution aerial drone images annotated for adult and calf classifications.ResultsSEMI-DETR achieved a mean Average Precision (mAP) of 19.2 at 0.5:0.95 using only 34 labeled images (1% annotation budget), representing a 20% relative improvement over the best-performing supervised baseline. The performance advantage peaked at +6.6 mAP points at the 5% labeling level. Semi-supervised learning particularly improved calf detection, achieving a 24.3% relative improvement at 1% labeling compared with 16.8% for adults. Overall, comparable detection performance could be achieved with a 5- to 10-fold reduction in annotation requirements without substantially compromising detection quality.DiscussionThese findings demonstrate that semi-supervised learning can substantially reduce annotation requirements for automated marine mammal detection, with particularly strong benefits for the detection of calves. Improved calf detection has direct implications for population assessment and the monitoring of reproductive success and recruitment in beluga populations. This study establishes an empirical benchmark for semi-supervised marine mammal detection and provides practical guidance for allocating limited annotation resources in wildlife monitoring and conservation applications.

Read on the original site

Open the publisher's page for the full experience

View original article