Detecting Mammals in UAV Images: Best Practices to address a substantially Imbalanced Dataset with Deep Learning (1806.11368v1)

Abstract: Knowledge over the number of animals in large wildlife reserves is a vital necessity for park rangers in their efforts to protect endangered species. Manual animal censuses are dangerous and expensive, hence Unmanned Aerial Vehicles (UAVs) with consumer level digital cameras are becoming a popular alternative tool to estimate livestock. Several works have been proposed that semi-automatically process UAV images to detect animals, of which some employ Convolutional Neural Networks (CNNs), a recent family of deep learning algorithms that proved very effective in object detection in large datasets from computer vision. However, the majority of works related to wildlife focuses only on small datasets (typically subsets of UAV campaigns), which might be detrimental when presented with the sheer scale of real study areas for large mammal census. Methods may yield thousands of false alarms in such cases. In this paper, we study how to scale CNNs to large wildlife census tasks and present a number of recommendations to train a CNN on a large UAV dataset. We further introduce novel evaluation protocols that are tailored to censuses and model suitability for subsequent human verification of detections. Using our recommendations, we are able to train a CNN reducing the number of false positives by an order of magnitude compared to previous state-of-the-art. Setting the requirements at 90% recall, our CNN allows to reduce the amount of data required for manual verification by three times, thus making it possible for rangers to screen all the data acquired efficiently and to detect almost all animals in the reserve automatically.

• The paper addresses the challenge of detecting mammals in imbalanced UAV image datasets using deep learning, proposing methods to significantly reduce false positives while maintaining high recall.
• Key strategies include class-weighting, curriculum learning, hard negative mining, introducing a border class, and carefully timed data augmentation to improve CNN performance on imbalanced wildlife image data.
• The approach achieved a dramatic reduction in false positives (from thousands to hundreds at 90% recall) compared to previous methods, making large-scale wildlife censuses via UAVs more practical and efficient.

Detecting Mammals in UAV Images: Best Practices for Tackling Imbalanced Datasets with Deep Learning

The effective monitoring of wildlife populations is essential for conservation efforts, particularly for endangered species. Traditionally, these tasks involved manual surveying techniques, which are both hazardous and costly. Recent advancements in UAV technology present a promising alternative, allowing for safer and more cost-efficient surveys. The paper by Kellenberger et al. focuses on using convolutional neural networks (CNNs) to automate the detection of large mammals in UAV images. This approach confronts the challenge of processing a substantially imbalanced dataset, where the number of background samples vastly exceeds the number of detectable animals.

The authors address a critical issue in applying CNNs to wildlife monitoring: the overwhelming imbalance between the negative (background) and positive (animal) samples in large real-world datasets. They propose a method that significantly reduces the number of false positives without sacrificing the detection rate. Key recommendations are made to overcome this imbalance, utilizing CNNs' capacity for deep feature learning optimized for image-specific challenges such as diverse animal appearances and environmental changes.

Methodological Innovations

The paper presents a variety of strategies to effectively train CNNs on imbalanced datasets. Among the most significant contributions are:

1. Class-weighting and Curriculum Learning: These methods are employed to mitigate the influence of the dominant background class over training. Class-weighting involves assigning higher importance to errors in the less frequent classes (animals) to guide the learning process. Curriculum learning starts the training on a balanced subset of the dataset before scaling up to the imbalanced whole, allowing the network to better anchor its learning of minority classes.
2. Hard Negative Mining: This technique focuses on locating and correcting the highest confident false positives, which helps improve model precision. By continuously refining the model to reduce such errors, the balance between detecting true positives and minimizing false positives is improved.
3. Border Class Introduction: This strategy deals with the problem where adjacent animal and background pixels can confuse the model due to overlapping receptive fields. A border class helps the CNN differentiate areas around animals, reducing confusion and subsequent false alarms.
4. Data Augmentation: Advanced augmentation strategies, particularly rotational augmentation, are applied progressively in the training cycle. While augmentation aids in making the model robust to different viewing angles and animal orientations, findings reveal that its timing in the training process is crucial for it to be effective without causing confusion.

Results and Evaluation

The CNN approach proposed in the paper not only effectively minimizes false alarms but also maintains high recall rates, critically outperforming existing methodologies. For a recall rate of 90%, the CNN reduced false positives from thousands, as seen in previous methods, to mere hundreds. This dramatic improvement underscores the effectiveness of the new strategies in handling imbalanced datasets over large census areas like the African savanna.

The authors introduce a robust evaluation protocol that prioritizes the practical objectives of wildlife censuses. This involves focusing on counting rather than pinpointing animal locations, reflecting fieldwork realities where core goals are accurate population assessments rather than spatial precision.

Implications and Future Directions

The paper's implications extend beyond animal censuses. The proposed methods to handle imbalanced datasets can influence a broad array of image recognition tasks in similarly challenging conditions. For wildlife monitoring, the reduction in verification effort facilitated by these advances means that large-scale UAV surveillance is more practicable and less resource-intensive.

Future work might explore integrating more sophisticated data generation methods, like generative adversarial networks, to further alleviate class imbalance issues. Additionally, adapting these methodologies for other ecological monitoring tasks where imbalance and class diversity present challenges could be valuable.

In conclusion, the paper offers a notable contribution to wildlife monitoring using UAVs, proposing well-founded enhancements to CNN training protocols that address imbalanced datasets' inherent challenges. The methodologies outlined herein enhance the efficiency of existing systems significantly, offering a sound pathway to improved wildlife conservation efforts.