Using Very High-Res Satellite Imagery and Deep Learning/Machine Learning to detect African Elephants
TLDRThe project, part of a PhD at the University of Oxford, explores the use of high-resolution satellite imagery and deep learning to detect African elephants. Collaborating with experts in zoology, computer science, and geospatial units, the team aimed to address the rapid species loss during the sixth mass extinction, largely due to human presence. The study built upon previous methods that used satellite imagery to monitor wildlife in open landscapes like the Arctic. The challenge was to adapt these methods for elephants, which move between various habitats. Using TensorFlow's object detection API, the team trained a model to identify elephants in satellite images, achieving comparable accuracy to human annotators. The study's success demonstrates the potential for satellite imagery and deep learning to contribute to wildlife conservation efforts, with future improvements expected from new satellite constellations offering higher resolution and revisit capabilities.
Takeaways
- 🎓 The project is part of a PhD research at the University of Oxford's Department of Zoology, aiming to use high-resolution satellite imagery and deep learning to detect African elephants.
- 🤝 The collaboration includes experts from the University of Oxford, the University of Bath, and the University of Twente, combining expertise in zoology, machine learning, and geospatial analysis.
- 🌍 The motivation is driven by the sixth mass extinction, where human presence is a significant factor in the rapid loss of species, and the need to monitor wildlife populations for effective conservation strategies.
- 🐘 The challenge is to monitor elephants, which move between habitats and are not as easily visible in satellite imagery as species in more open environments like Arctic penguins or marine mammals.
- 📈 Existing studies have demonstrated the potential of satellite imagery to monitor wildlife in open landscapes, but elephants present a unique case due to their size, movement, and habitat preferences.
- 🛰️ The project utilized 11 high-resolution images from the WorldView 3 and 4 satellites covering different areas and seasons in South Africa's Elephant Park.
- 🖼️ Image processing involved pan sharpening to achieve a color image at 31-centimeter resolution, which was then annotated to create a dataset for training the machine learning model.
- 🤖 A TensorFlow Object Detection API was used to build a model for automated detection, which was compared to human annotations for accuracy.
- 🏆 The Convolutional Neural Network (CNN) performed comparably to human annotators and was able to generalize detection to a different geographical area and satellite resolution in Kenya.
- 🚀 The study demonstrates the potential for using satellite imagery and deep learning for wildlife monitoring, although limitations include cost, revisit time, and tasking capabilities.
- 🌟 Upcoming satellite constellations with higher resolution and revisit frequency, such as WorldView Legion and Pleiades Neo, could significantly enhance the ability to monitor wildlife populations.
Q & A
What is the main objective of the project described in the transcript?
-The main objective of the project is to use very high-resolution satellite imagery and deep learning to detect African elephants, with the aim of conducting a census survey and monitoring population trends for effective conservation efforts.
Who are the key collaborators involved in this project?
-The key collaborators include the speaker who is a PhD student in the Department of Zoology at the University of Oxford, Olga, a machine learning expert formerly at Oxford and later at the University of Bath, Professor David Macdonald, Tj Wang from the Geospatial Unit at the University of Twente, and Steven Reese from the Department of Machine Learning at Oxford.
What are the challenges faced when using satellite imagery to monitor wildlife, particularly elephants?
-Challenges include the cost of high-resolution satellite imagery, cloud cover, the need for revisit time and tasking capabilities, the ability to distinguish between different species and individuals, and the difficulty of monitoring wildlife in areas with dense vegetation or closed canopy.
How does the machine learning model used in the project work?
-The machine learning model used is a Faster Region-based Convolutional Neural Network (Faster R-CNN) that processes the images to output bounding boxes around objects of interest and classifies each of these bounding boxes. It operates in two stages within one coherent architecture: first, it uses convolutional layers to extract features from the input image, then it proposes rough bounding boxes where objects might be located, and finally, it refines these boxes and classifies the objects within them.
What are the advantages of using satellite imagery for wildlife monitoring?
-The advantages include the ability to cover very large areas, monitor cross-border regions without needing aviation permissions, conduct repeat surveys at short intervals, and avoid disturbing animals due to the unobtrusive nature of the technique. It also overcomes logistical challenges associated with setting up camera traps or manned aircraft surveys.
How was the performance of the machine learning model evaluated?
-The performance was evaluated using the F2 score, which is a measure that gives more weight to false negatives. The model's results were compared with human annotation performance, where a group of volunteers labeled the same satellite images to provide a baseline for comparison.
What are the future prospects of using satellite imagery for wildlife monitoring?
-The future prospects are promising with the launch of new satellites like the WorldView Legion constellation and the Pleiades Neo constellation, which will provide higher resolution imagery and more frequent revisit times. This will allow for more efficient monitoring of different elephant populations and potentially other species that aggregate in large groups.
How was the issue of double counting elephants addressed during the study?
-Double counting was mitigated by having the speaker label the images, followed by a double-check from Olga. Additionally, a separate group of volunteers also labeled the images, and the counts from human annotators were compared with the counts from the convolutional neural network to ensure accuracy.
What role did the European Space Agency play in facilitating this research?
-The European Space Agency provided a grant to researchers, which allowed them to pay for third-party imagery. They facilitated the tasking of imagery from Mali through Maxar, which was used in the research.
How did the researchers deal with the challenge of small object detection in satellite imagery?
-To deal with small object detection, the researchers used a master student's preliminary results that suggested enlarging the images through interpolation could help improve the detection of small objects like elephants in the images.
What other species or environmental factors were considered for monitoring using satellite imagery?
-While the primary focus was on African elephants, the researchers are actively seeking to obtain imagery for other species like rhinos. They are also considering the possibility of using satellite imagery to detect environmental factors such as deforestation, wildfires, and floods.
Outlines
📚 Introduction to the Elephant Monitoring Project
The speaker introduces a project that is part of their PhD at the University of Oxford's Department of Zoology. They discuss the interdisciplinary collaboration with a former postdoc researcher, Olga, who is now at the University of Bath. The project aims to use high-resolution satellite imagery and deep learning to detect African elephants, addressing the rapid rate of species loss during the sixth mass extinction. The project is conducted in collaboration with multiple supervisors and aims to monitor elephant populations to inform conservation efforts.
🌐 Satellite Imagery and Machine Learning for Wildlife Conservation
The speaker elaborates on the use of satellite imagery to monitor wildlife, highlighting previous studies in various environments like the Arctic and seascapes. They discuss the advantages of satellite monitoring, such as covering large areas and avoiding human disturbance to wildlife. However, they also mention challenges like cloud cover, high costs, and the need for high-resolution imagery. The speaker then focuses on their work with African savannah elephants in South Africa's Elephant Park, noting the park's varied landscapes and the availability of satellite images for the area.
🖼️ Image Processing and Object Detection
The speaker details the process of using satellite images for elephant detection. They describe the image processing techniques, including pan sharpening to achieve a 31-centimeter resolution. The process involves manually labeling images to identify elephant locations and then using TensorFlow's object detection API to build a model for automated detection. The model's performance is compared with human annotations, with the goal of achieving comparable accuracy.
🐘 Elephant Detection Challenges and Results
The speaker discusses the challenges of detecting elephants in satellite images, particularly when they blend with the background or are in forested areas. They present results from testing the machine learning model, showing that it can accurately detect elephants in both homogeneous and heterogeneous backgrounds. The model's performance is measured using the F2 score, which balances precision and recall, with a focus on minimizing false negatives.
📈 Performance Comparison and Generalization of the Model
The speaker provides a comparative analysis of the machine's detection performance with human observers. They highlight that the machine's results were mostly comparable to human performance and even outperformed humans in detecting elephants in heterogeneous areas. The speaker also mentions testing the model with an image from Kenya, which was not part of the training data, demonstrating the model's ability to generalize to new areas and detect different sized elephants, including calves.
🚀 Future Prospects and Limitations
The speaker concludes with a discussion on the future of satellite imagery for wildlife monitoring. They mention upcoming satellite constellations that will provide higher resolution and more frequent imagery, potentially allowing for monitoring of more species. However, they also acknowledge limitations such as cost, revisit time, and tasking capabilities. The speaker expresses optimism about the role of deep learning in handling the vast amount of imagery and enhancing wildlife monitoring methods.
🌴 Challenges of Tropics and Closed Canopy
The speaker addresses questions about the feasibility of using satellite imagery in tropical regions with dense vegetation. They mention that while the study did not directly address this, there is potential for oblique imagery to peer under canopies. The speaker also discusses the limitations of counting elephants without double counting and the importance of having non-overlapping images taken at different times.
🔍 Expanding the Study to Other Species
The speaker expresses interest in expanding the study to include other species, such as rhinos, but notes the challenge of obtaining suitable imagery for training the model. They also discuss the potential of using imagery from sources like Planet Labs for detecting large aggregations of animals or environmental proxies for wildlife presence.
🛠️ Image Sampling and Model Exploration
The speaker talks about the tools used for image sampling and the exploration of different machine learning models. They mention the use of an open-source tool created by the computer vision group at Oxford for image annotation. The speaker also discusses the preliminary results of using image interpolation to increase the size of objects in the images, which improved detection accuracy.
Mindmap
Keywords
💡High Resolution Satellite Imagery
💡Deep Learning
💡Machine Learning
💡African Elephants
💡Wildlife Conservation
💡PhD Research
💡Species Decline
💡Population Counts
💡CNN (Convolutional Neural Network)
💡Object Detection
💡Species Monitoring
Highlights
A PhD project at the University of Oxford is utilizing high-resolution satellite imagery and deep learning to detect African Elephants.
The collaboration includes experts from the Department of Zoology, Department of Computer Science, and the Geospatial Unit at the University of Twente.
The project aims to address the rapid rate of species loss during the sixth mass extinction, largely driven by human presence.
Satellite imagery has been used to monitor wildlife in open landscapes like the Arctic; however, African Elephants present a challenging case due to their movement between habitats.
The study explores the feasibility of conducting a census survey for African Elephants using satellite imagery, which has not been tested before.
The research used 11 high-resolution satellite images from the World View 3 and 4 satellites, covering both homogeneous and heterogeneous areas of a South African park.
Images were processed using TensorFlow's object detection API, with a focus on accuracy over speed.
The machine learning model was trained to detect elephants with a high degree of accuracy, comparable to human annotation.
The study found that the deep learning model was able to generalize and detect elephants in different geographical areas and various resolutions of satellite imagery.
The research highlights the potential of using satellite imagery for wildlife monitoring, despite challenges such as cloud cover and cost.
The study suggests that upcoming satellite constellations with higher resolution and revisit times will make monitoring more viable.
The research is a proof of concept that satellite imagery combined with deep learning can be used for wildlife detection and conservation efforts.
The study did not provide a total population count but demonstrated the potential for counting elephants in different environmental contexts.
The project faced limitations such as the inability to see through dense forest canopies and the high cost of high-resolution imagery.
The research team is actively seeking to expand the detection to other species, such as rhinos, and to differentiate between species using the model.
The study explored the idea of artificially increasing the size of objects in images to improve detection rates by deep learning models.
The research demonstrates the potential of using satellite imagery for large-scale wildlife monitoring, which could transform conservation strategies.