Machine Learning for Bio-diversity Monitoring and Tracking at Scale
Project Summary
The need to protect the environment has taken center stage in recent years as the damage to the Earth’s natural ecological balance has become increasingly tangible. Natural reserves are set to protect the biodiversity. With this good intent comes great challenges. Understanding the populations of animals and vegetation living inside the reserve is a key parameter to secure the equilibrium and the continuity of the ecosystem. We propose to investigate the problem of biodiversity monitoring and tracking using machine learning techniques for mining useful information from complex data, image/videos.
Deep learning methods have proven to be efficient in the image processing field. In fact, their introduction in this area has reshaped the way images are processed and understood. At the same time, this raised expectations with regards to performing and automating the processing and understanding of images, along with understanding their content. A large number of works have been performed and many methods have been proposed to tackle different aspects of the problem. The novelty in this project resides in two folds: it is the first time that the performances of existing algorithms will be evaluated in such large magnitude and diversity of experiments, which will support the investigators to propose new algorithms. The second novelty resides in the application of the approach in the area of biodiversity tracking and monitoring in the desert. In fact, the investigators are not aware of any existing efforts that suggest the application of image mining in the desert.
The investigators intend to work closely with the DDCR to help detect and monitor the different species, their evolution, movement, and gender. This will directly help the DDCR in better estimating the number of animals, predict their evolution, understand their moves, and understand the density of those animals. The project is expected to be beneficial for the wildlife and the biodiversity.
Research Objectives
- Collect a significant quantity of data to test a number of existing methods in real-life situations.
- Determine benchmarks to be used in the evaluation of future methods.
- Setup optimized techniques for scanning large desert areas for resources management.
- Propose methods for identifying animals, their family, gender, and movements.
- Create a set of methods and workflows to implement the findings of this project.
Methodology
This research will make use of modern image mining techniques to contribute to the area of environmental protection through automatic animals’ recognition and identification in wild areas. In this project, we will begin by creating a large scale and controlled experiments protocol to evaluate the top existing methods in research w.r.t. our context, i.e., animals identification in wild areas. We will then follow an experiment-based methodology in which real situations and real datasets (e.g., UAV captured images and videos) will be used for observation and learning models. These datasets will serve as a foundation for enabling the extraction of actionable knowledge. Once the experiments are performed; we will proceed to researching a method (or methods) that would integrate the knowledge and the learnings from the experimental phase.
At a high level, we will follow the steps shown in the next section.
- Data collection and preparation: As a backbone of our research and experiments, the data to analyze will be related to litter in wild areas, i.e., desert in our case. As UAVs will be largely used here, the configuration of the experiments must be taken care of to ensure an objective assessment and interpretable quality of the outcomes. The datasets have to be tagged and classified to make sure that we can build the expected models. Considering the nature of the environment, different attempts may be considered to ensure a high quality of data is gotten.
- Benchmarking of existing methods: Relying on the collected data, a large experimental setup will be put in place to evaluate the most recent and advanced techniques available in the literature to identify animals, as well as movements and related concepts.
- Model(s) development: The objective here is to elaborate new methods for supporting animals identification in wild areas. While developing the models, we should take into consideration that we are dealing with real-time and heterogeneous data taken under different constraints.
- Implementation and Evaluation: This step will basically run across the entire project. Its objective will be to define the quality of the proposed methods and approaches in terms of effectiveness and performance. We will perform this objective through formal methods and subjective evaluations.
Expected Outcomes
- The project will provide a set of data, described in detail and capturing different situations and constraints.
- This report will contain a rigorous description of a set of experimental protocols that will govern the test of different methods and techniques in the literature.
- This deliverable is the final outcome and expectation of the project. It is a software platform that is dedicated for end-users and which will allow them to analyze real-time streams coming from unmanned aerial vehicles. It will embed artificial intelligence techniques that will not only reduce, e.g., the noise in image streams but also the selection of the best configuration to use by the UAV depending on the external constraints. The software is expected at the end of the project in its final version.
- We will target at least 2 top conferences/journals publications. These publications will be expected during the second year of the project, mainly.
- This report is a classical final report. It will be provided at the end of the project and will describe the achieved objectives and outcomes of the project.
