VN September 2025

Vetnuus | September 2025 33 This suggests that Positive Certainty is finding less informative middens than MultimodAL, which is biased towards selecting the warmest middens. A comparison of Figures 7 and 8 indicates a positive relationship between the accuracy achieved by the active learning system and the number of positive samples it is able to find. In addition, the fact that both versions of MultimodAL find roughly the same number of middens throughout the active learning process, despite MultimodAL: T + R achieving higher accuracy at the beginning, is further evidence for it finding more informative middens than MultimodAL: Thermal+RGB Fused. Figure 8 also demonstrates that MultimodAL is successfully addressing the class imbalance issue in our dataset. Both versions of MultimodAL find over 65% of the middens by 500 images labelled. This is quite remarkable given that only 71 out of the 9,736 images in the training set are middens. Hence, although less than 1% of the images given to the active learning system are middens, our active learning algorithm is able to discover 46 of them by the time 500 labels have been requested. This 9% positivity rate is 12x larger than the overall positivity rate of middens in the training set. Perhaps more impressively, after just 100 queries, MultimodAL has found over 34% of all the middens, whereas none of the baselines manage to find more than 8%. The number of middens found is significant because these are confirmed by humans, so we are highly confident that these are indeed middens. Once the labelling budget has been exhausted, the remaining unlabelled images are classified using the trained model, but because the model is not 100% accurate, we have more uncertainty as to whether the images classified as middens are truly middens. By achieving competitive accuracy and quickly identifying middens in an unlabelled dataset, MultimodAL drastically alleviates the labelling burden, as quantified in Table 2. If we assume each image takes 30 seconds for a domain expert to label, then labelling every image in the dataset would require over 80 hours, but labelling 500 images would take under 4.2 hours, which is a significant amount of time saved in a resource-constrained domain — all while identifying 65% of all the middens in the dataset. This demonstrates the power of an active learning system that employs a domain-inspired ranking technique to find rare positive samples in a highly imbalanced dataset. Midden Mapping From the locations of the middens in the site, we create the first-ever midden map, presented in Figure 9, and extract actionable insights. We emphasise that we are not showing landscape features on our map to protect rhino territories from poachers. Because rhino middens are used for territorial marking, by mapping these middens, we can also figure out where the rhinos’ territories are. We can then target ranger patrols and ecological monitors at the middens to better protect the rhinos from poachers. Unexpectedly, the midden map reveals that the middens are not distributed randomly across the region but instead form clusters. Using K-means, we have identified six such clusters on our site. We recommend that antipoaching efforts prioritise those areas of high midden density to increase their efficacy. Furthermore, we find that the rhino middens are often located along animal paths. Ecological monitors can thus focus their efforts along these paths since they are likely being used by rhinos. Additionally, knowing where the middens are is useful because monitors can visit them to more quickly find rhinos or check for fresh dung as a way of confirming the presence of rhinos in the area. This is crucial to demonstrating that the rhinos are being effectively managed and that the area should continue to be funded and resourced. Thus, by more effectively targeting scarce resources, we expect MultimodAL to have a significant impact on rhino conservation. Future Work In advancing a decade-long collaboration with South African National Parks, we will use MultimodAL to map middens across other sites in Kruger National Park. To do this, we will use a model pre-trained on the labelled dataset studied herein to warm-start an active learning system, further lessening the labelling burden. Unique terrain features, such as termite mounds present in some other sites, could increase the advantages of multimodality. To support this work, we hope to develop an interactive user interface for MultimodAL to allow domain experts to label images queried by the model(s) without needing to work directly with computer scientists. We also present two ideas for building on the MultimodAL system. First, the algorithm could be extended by dynamically weighting the maximum thermal pixel values and the model(s)’ output. Second, we could define the midden detection task as one of image segmentation rather than classification and use the model from Kirillov et al. [2023]. We also encourage the evaluation of the MultimodAL system on imagery from other domains (e.g., wildfire or animal detection). To facilitate this future work, we make our code publicly available at https://github.com/lgordon99/rhino-middendetector. In contributing to the UN’s Fifteenth Sustainable Development Goal, we encourage protected area managers to use the MultimodAL system to map rhino middens across their landscapes. However, we must keep our dataset private due to its sensitive nature with respect to the rhino poaching crisis. In accordance with the “Leave no one behind” principle [UN, 2023a], we believe rhino midden maps can empower park rangers and ecological monitors in carrying out their indispensable work in conserving the planet’s last wild places. Conclusion We make several contributions in this paper. In collaboration with South African National Parks, we map rhino middens for the first time and display where anti-poaching patrols ought to focus their efforts. Thus, we have demonstrated a way of finding and protecting rhinos without directly tracking them. We find that models passively trained on thermal, thermal+RGB fused, or thermal+LiDAR fused imagery achieve high accuracy in distinguishing between midden and empty images. To alleviate the labelling burden intrinsic to passive learning, we introduce MultimodAL, a novel active learning methodology for highly imbalanced datasets that is applicable when the instances can be ranked according to an informative metric. In addition to exploiting domain knowledge, this method also accommodates multimodal data. We demonstrate that the multi-model instantiation of the algorithm outperforms a single-model system on our dataset for a small labelling budget. Overall, both MultimodAL systems outperform baseline query strategies and find more middens than the baselines for up to 400 labels provided. Furthermore, only MultimodAL achieves an accuracy statistically indistinguishable from that of the best passive model with 94% fewer labels, saving domain experts dozens of hours in data annotation efforts. By introducing this scalable method for mapping middens, we present new insights to support conservation practitioners in advancing rhino conservation across southern Africa. v Acknowledgments Research was sponsored by the DARPA and was accomplished under Grant Number: HR001122C0182. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of DARPA or the U.S. Government. The U.S. Government is authorised to reproduce and distribute reprints for Government purposes, notwithstanding any copyright notation herein. J.A.K. was supported by an NSF Graduate Research Fellowship under grant DGE1745303. South African National Parks is thanked for logistical and scientific support, as well as permission to work in Kruger National Park. All drone flights were performed with approval from the South African Civil Aviation Authority and South African National Parks Airwing. Mode Test Accuracy Labelling Budget Labelling Time Passive 0.86 ± 0.01 9,736 81 hours Active 0.84 ± 0.01 500 250 mins Active 0.81 ± 0.01 200 100 mins Active 0.73 ± 0.01 50 25 mins Table 2: Comparison of passive (Thermal+RGB Fused) and active (MultimodAL: Thermal + RGB) test accuracy, labelling budget, and labelling time Figure 9: Midden map for a 2x2-km site in Kruger National Park. Blue dots are middens. Red stars are centres of midden clusters References available on request.

RkJQdWJsaXNoZXIy OTc5MDU=