Habitat fragmentation occurs when large animal habitats are broken up into smaller areas, such as a road being built through a forest. This breaking up of habitats can impact biodiversity by preventing the movement of certain species between habitats. However, how this habitat fragmentation occurs and the effects it has on biodiversity are not fully understood and have been widely debated.
Current methods for calculating fragmentation metrics exist in the field of landscape ecology and involve the calculation of landscape or patch metrics from land cover maps in order to examine ecosystem fragmentation. However, these methods are limited in their ability to answer important questions relating to how connected certain habitats are and how they evolve over time.
As a group of MRes students at Cambridge University, we were tasked with using machine learning in combination with remote sensing data to address issues relating to biodiversity loss and decided to tackle this problem by producing a new package named GeoGraph.
GeoGraph proposes a graph-based framework for assessing habitat fragmentation. It takes land cover data as an input and converts each patch of a single land cover type into a geospatially referenced graph node. Edges are added to the graph to represent the proximity between adjacent land cover patches. Once the geospatial graph has been created, users can input a list of land cover classes that they define as part of the same animal habitat. These combined classes then become ‘habitat components’ in the graph framework. Edges can then be added between the habitat components that an animal living in the habitat can be expected to travel between. This travel distance is also a parameter that can be adjusted by the user for the study of a particular animal.
This graph-based framework provides many new implications for the assessment of habitat fragmentation. Firstly, it provides a very flexible and user-friendly framework in which users can input their own land cover data and define their habitats and species of interest. Secondly, it comes with the ability to calculate novel metrics for the connectivity of habitats that may be useful for ecologists studying the effects of habitat fragmentation on certain animal species. Also, it allows for more rigorous analysis of how certain habitats are changing over time. Since the graph nodes are geospatially referenced, it is easier to define the change in specific habitats over time as opposed to the change in an entire land cover type over a given area as done by previous methods.
Through its interactive user interface, GeoGraph can also provide important insights to policymakers. The area around a habitat component in which a certain animal may have the ability to travel can be visualised on a map. Through this visualisation potential new conservation regions can be suggested by overlapping areas. It is also possible to visualise the areas that are most at risk from further habitat fragmentation by highlighting habitat components that only have a single connection to another component. This has implications for conservation policy, as policymakers can ensure that these at risk areas do not become more fragmented.
The main limitation of GeoGraph for the assessment of habitat fragmentation is its reliance on suitable land cover products. Often the freely available land cover products are too coarse in resolution to be useful for assessing habitat fragmentation. Alternatively, land cover products are often not formulated with ecosystem assessments in mind, so they may not contain ecologically relevant classes. This may limit the potential for GeoGraph to be applied at scale in certain areas.
We attempted to address this problem in our work by formulating a land cover classifier using a convolutional neural network trained on satellite imagery and a manually labelled habitat dataset. The main difficulty we encountered was that classes that are ecologically significant for habitat assessments may appear very similar in the satellite imagery so the classifier we built was not effective in distinguishing them. This is an area in which further work is needed in order to build an end-to-end product to assess habitat fragmentation in any area through inputting a satellite image, producing a land cover map, calculating the geograph and then performing the analysis.
Currently, GeoGraph provides a new framework for assessing habitat fragmentation from land cover maps and has the potential to improve ecologists’ understanding of habitat connectivity, how this evolves over time and how it affects biodiversity. Through effective visualisations, it also provides an important tool for policymakers in informing key areas in need of protection to prevent further habitat fragmentation.
- For more details see https://github.com/ai4er-cdt/gtc-biodiversity.
- A Binder demo can be found at https://bit.ly/geograph-binder.
- Project documentation is at https://geograph.readthedocs.io/.
Did you miss Climate Exp0? You can watch the highlight videos from all five days of the conference here. All conference content and sessions will be freely and permanently available on Cambridge Open Engage from mid-June. For those who attended Climate Exp0, you can view all sessions from Climate Exp0 now via the Climate Exp0 homepage, using your conference login.