Conservation in action: Cost-effective UAVs and real-time detection of the globally threatened swamp deer (Rucervus duvaucelii)

Uncrewed aerial vehicles (UAVs) have proven to be successful tools for ecological monitoring, providing excellent visual resolution and the ability to cover large areas with spatial accuracy. Artificial Intelligence has further improved the capabilities of UAVs vision through object detection. While deep learning has shown significant success in pattern recognition, it still faces challenges in real-world scenarios. In this study, we focused on enhancing the potential of UAVs for wildlife detection and monitoring, specifically focusing on the globally threatened swamp deer (Rucervus duvaucelii). To improve the accuracy of animal recognition with UAVs, we used single-stage detectors YOLO (You Only Look Once) V3, V5, V7, V8, Object detection V3 and DETR (DEtection TRansformer). We trained our model using 48,957 augmented images derived from a dataset of 8210 original true dataset. The result shows the superior performance of Object Detection 3.0 and YOLO V8, achieving a precision score of more than 92 % and a F1 score of more than 85 %, compared to DETR, YOLO V7, V5, and V3. Overall, our study provides an efficient, cost-effective, and accurate detection framework for ecological monitoring that can be non-invasively and least distractively used in various demographic studies of cervids in different regions and habitat types. We have also developed a UAV-based real-time object detection framework that seamlessly integrates with front-end technologies, enabling live detection results regardless of connectivity. This framework operates on a local server, synchronizing with consumer-grade UAVs at a rate of 32 frames per second (fps) with 320 pixel resolution using a frame sampling technique, notably without requiring a dedicated Graphical Processing Unit (GPU). This deliberate choice of not using GPU underscores the commitment to cost-effectiveness and aligns with the research's purpose, prioritizing accessibility and affordability for broader scientific exploration. A least distractive ecological sampling technique was optimized and a maximum of 77 deer were detected and counted in real time within Haiderpur wetland in the Hastinapur Wildlife Sanctuary. This methodology can be replicated and fine-tuned to study other threatened species of conservation priority. This study exemplifies how combining UAVs with deep learning can facilitate species monitoring and population count estimation and be adopted by forest managers to support conservation decisions. [Displa