Remote Sensing of Nature

Measuring the world's forest carbon and biodiversity is made possible by remote sensing instruments, ranging from satellites in space (Landsat, Sentinel, GEDI) to ground-based sensors (ecoacoustics, camera traps, moisture sensors) that take regular samples and are processed into time-series metrics and actionable insights for conservation and human development. However, the algorithms for processing this data are challenging as the data is highly multimodal (multispectral, hyperspectral, synthetic aperture radar, or lidar), often sparsely sampled spatially, and not in a continuous time series. I work on various algorithms and software and hardware systems we are developing to improve the datasets we have about the surface of the earth.

Mapping nature on earth

Figuring out where things live on the planet's surface from satellites requires a lot of data processing, and tricks to work around the fact that we can't easily see through clouds (when using optical sensors) or handle very sloped surfaces (if using lidar) or peek through the top of a dense forest canopy (especially in tropical forests). Along with colleagues in the Cambridge Centre for Earth Observation and especially David A Coomes, I've been working on a few projects that aim to improve the quality of the data we have about the surface of the earth.

The main research question we're tackling is how to improve our knowledge about where most wild species live on the planet, so that we can better protect their receding habitats. And in particular, our knowledge of where rare plant species' live is surprisingly data deficient.

Satellite and drone sensing

Old-growth tropical trees have the big advantage of being relatively easily visible from the air, and we've been developing a robust satellite and drone processing pipeline as part of the Planetary Computing project. James G. C. Ball and Sadiq Jaffer have leading an effort to use this data to develop a new approach for mapping tropical tree species. They link a multi-temporal implementation of a CNN method to segment tropical forest tree-crowns from aerial photographs, to ML classifiers that can identify species from hyperspectral data.

Read more about it in the "Harnessing temporal & spectral dimensionality to identify individual trees in tropical forests" preprint.

Common base maps for Area of Habitats

AoH calculations per species are really important to agree on, and are generated from a combination of range maps, habitat preferences, climatic variables and occurrence data. Michael Dales and I are working with other developers of biodiversity metrics (such as IUCN's STAR team) which also require AoH maps to develop a common base layer that can be maintained communally. This will also make it far easier to pinpoint algorithmic differences between STAR and LIFE rather than simply varying because of differring input data. You can find the code for our area-of-habitat calculators for 30k terrestrial vertebrates online, and (thanks to a UKRI funded project in 2024) this will be expanded this to include plants.

Species Distribution Modelling

One use for AoH maps is to turn them into Species Distribution Models, which is a way to predict where species are likely to be found based on environmental variables and occurrence data. Emily Morris worked on a new method that uses a combination of satellite data and machine learning to predict the distribution of species across the globe, with her focus being on proteas. Read more about it in Towards Scalable Deep Species Distribution Modelling using Global Remote Sensing.

Ground-based sensing with the Terracorder

In 2024, I started collaborating with Josh Millar over at Imperial College on developing a low-cost sensor device designed for long-term deployment in remote nature areas as well as urban environments. Since in-situ sensing devices need to be deployed in remote environments for long periods of time, minimizing their power consumption is vital for maximising both their operational lifetime and coverage. We started from an ESP32 base (due to the lovely 16-bit ultra-low power mode) and have been prototyping the "Terracorder" as a versatile multi-sensor device. Read more about it in Terracorder: Sense Long and Prosper.

Since I've been exploring spatial networking with Ryan Gibb (see Where on Earth is the Spatial Name System?), we've also been figuring out whether a combination of reinforcement learning and spatial networking knowledge might take this device to the next level of usability. We've been experimenting with using an on-device reinforcement learning scheduler. When evaluating our prototype scheduler against a number of fixed schedules; the scheduler captures more than 80% of events at less than 50% of the number of activations of the best-performing fixed schedule. We're currently working on a collaborative scheduler can maximise the useful operation of a network of these Terracorders, improving overall network power consumption and robustness.

Applications to human health

Ultimately, it would also be nice to understand the impact of more natural spaces on human health as well. After all, we not only need to protect unspoilt nature, but also need to make sure that highly urbanised areas are also liveable. Andres Zuñiga-Gonzalez, Ronita Bardhan and I have been investigating the impact of green spaces in cities. These have been demonstrated to offer multiple benefits to their inhabitants, including cleaner air, shade in sunny periods, and a place that contributes to mental well-being. In addition, trees in cities are home to several species of animals and work as a nature-based solution that can sequester CO2 and regulate water storage in urban ecosystems.

So far, we've been working on using a combination of remote sensing data and local metrics to connect the dots about the impact of urban green spaces on human health. Read more about our work in "Green Urban Equity: Analyzing the 3-30-300 Rule in UK Cities and Its Socioeconomic Implications" and the project in The role of urban vegetation in human health.