Species Distribution Modelling

As in previous weeks, I’ve spent a good amount of time playing around with species distribution modelling with TESSERA.¹ In particular, I’ve been using the GeoPlant dataset used for the annual GeoLifeClef competition. Over the last month or so I’ve made a big mess with a myriad of scattered experiments. I’m now attempting to bring everything together with a final “grid” of models that encapsulate my various conclusions. I’ll run this grid over the coming period.

As a first taster, here is a bar chart showing the performance of my “fiducial” model² with three input data modalities: Sentinel-2 images, AlphaEarth embedding “images”, and TESSERA v1.0 embedding “images”. All images are 128x128 10m pixels. Performance is measured by the sample-weighted F1 score used in the GeoLifeClef competition.

The two geospatial foundation models, AlphaEarth and TESSERA, beat the “traditional” remote sensing modality (Sentinel-2) by a convincing margin. Of the two, TESSERA just about inches ahead of AlphaEarth.³

Bird Assemblage Mapping

These past few weeks I’ve spent one day per week working on bird-derived habitat mapping project I described in my previous weeknote.⁴ I haven’t gotten as far as making a map yet, as I’ve been spending quite a bit of time understanding and preparing the data that have been shared with me by my collaborators.

The main dataset is a table of associations between bird species and the habitat types under the Habitats of the World typology. To give a sense of what this looks like, the bar chart below shows the number of species in each of the 57 African habitats, disaggregated by association strength.⁵

In addition to this table, I was also given a wealth of birding checklists from various sites in South Africa and Australia. Here, a somewhat irksome problem creeps in: species identifiers. Many of the birdlists use a different name for a given species compared with the bird-habitat table.⁶ On Friday, over the course of several rather strong cups of coffee, I manually built a table of corrections.⁷ This was non-trivial: often, the name differences are not simply synonyms but taxon merges/splits. In the case of a split, I had to use geographic information to work out which of the derived species more likely corresponded to the observation. As part of my sanitisation, I moved to using the eBird “species code” as a more stable unique identifier for each species.

Painful, but actually quite fun! Plus I got to look at a lot of nice birds as I went along. For example, the Papuan pitta (Erythropitta macklotii… or is it Pitta erythrogaster??).

Crossword

Over a few sessions in July, I worked on a delightful crossword with my wife and some friends: the Guardian “Genius” crossword, number 277.

The Genius is a monthly crossword associated with a cash prize.⁸ The clues themselves are actually not too hard, but each puzzle does have an unusual twist which makes the overall puzzle quite a challenge. In #277, the rubric itself was hilariously cryptic:

The puzzle is a safe. Solvers must reveal its contents using four dials, applying cach to a quarter of the clues. Dial by dial in clue order, certain letters give a property of 23 cells to be highlighted:
Dial 1 clues: extra letter indicated by wordplay
Dial 2 clues: first letter of an extra word
Dial 3 clues: letter omitted by wordplay
Dial 4 clues (normal): clue’s first letter

I had to read this ten or twelve times before I understood it (/ gave up trying to understand it). The essence of it is that each clue in the crossword yields an additional letter, falling into one of four categories. Concatenate them in the right order to receive further instructions. Below is an image of the completed crossword, with colours indicating the four categories.

The additional letters, spelt out dial by dial, then say: CONTAINSAMELETTERASTODAYSCRYPTIC. Or, “contain same letter as today’s cryptic”. “Today’s cryptic” refers to the “normal” Guardian cryptic #30051, which was released on the same day as Genius #277. If one highlights just the cells in which both crosswords contain the same letter, one sees the hidden objects emerge:

Dagger, fake ID, diamond ring. These are the final answers which are submitted for the prize. I confess I can’t see a connection between the three (other than that they might go in a safe?), but I suspect there’s a cultural reference I’m missing.

In addition to a weird and wonderful puzzle concept, the crossword also had a new (to me) type of clue. 25 across:

____ ____ exemplify ruin? (7)

The answer is DESTROY (=“ruin”). The blank parts of the clue are “Does Troy” (there’s an extraneous letter O here, according to dial 1). So, the clue’s wordplay includes the literal answer!

Crochet

Last week was my wife’s and my “cotton” anniversary, and we exchanged gifts of cotton. For my part, I attempted my first ever three-dimensional crochet, a Crassula umbellata. Pattern from “Stylish Succulents to Crochet” by Sarah Abbondio. Photo below.

It was much harder than I thought it would be. I’m going to return to two dimensions for a while.

Synthetic Silviculture

A few weeks back, I met Dominik Michels from TU Darmstadt. He told me about some of the work his group is doing on forest simulations. This past week, I finally got round to reading some of their papers. For example: Synthetic Silviculture.³

The basic idea is that a real tree is represented by a topological tree: nodes and edges. It has various parameters that set its growth and branching rate. Neighbouring trees shade each other following ray physics calculations. Other processes (reproduction, fire, etc.) are also implemented.

Videos of the simulations seem really impressive. We talked a little about how we might use remote sensing to provide validation experiments for the simulations, and we also talked about conservation research questions one might address with such simulations, like simulating the effects of deer control in the Cairngorms.

15 Reasons for Conservation Monitoring

A few weeks ago, we had Prof. Hugh Possingham visit us to give a colloquium. His talk was, by design, rather provocative. In particular, he seemed to be advocating a philosophy where one should only pursue a nature monitoring project when one envisages a specific on-the-ground action that will result from the monitoring.

I found myself worrying quite a lot about this philosophy. While it seems pragmatic in principle, I fretted about all the data that would never have been collected had scientists always strictly followed this philosophy. For example, a great deal of knowledge we have today about effective conservation practice comes from conservationists continuing their monitoring schemes well into the post-intervention phase, even when they knew the outcomes of their interventions, when they had no further planned actions in mind, and when they might well have devoted their resources elsewhere.

This past week I went away and read his paper on the subject.⁴ The paper is actually much more circumspect than the talk. It lists 15 reasons for a conservation monitoring programme, framing them as a checklist that one can go through when planning a project. Included among the reasons are some of the more ‘nebulous’ reasons for data-gathering, like ‘basic research’, ‘taking the pulse of nature’, and ‘societal curiosity’. These are all framed as valid reasons to pursue research. The main point of the article is to encourage scientists to think a little more explicitly about their reasons to pursue different projects. Then, by having thought more explicitly, they can might be able to allocate their finite resources in the best way.⁵

I certainly feel I could do with taking this message on board! I’m not sure whether any of my current research can be classed as ‘conservation monitoring’. However, I might nonetheless consult this checklist next time I am planning a project, in order to be a little more clear and intentional about the expected impact of my research.

Leuven

Last weekend I visited Leuven, in Flanders. This was my first time visiting Leuven, and I thought it was delightful. It reminded me a little of Cambridge: a beautiful university town with narrow streets, many bicycles, and a winding river lined with willows and ancient buildings.

A highlight of the city is the UNESCO-listed “Great Beguinage”.⁶ The article I read beforehand recommended seeing some of the other beguinages of Flanders so that I could really appreciate the greatness of the Great Beguinage. It didn’t, however, explain what a beguinage was, so I went in blind.

As it turns out, a beguinage is a built community created by beguines: religious women who lived a communal life, without taking the strict vows of the convent. There are a number of these beguinages around the Low Countries, Flanders in particular. The Great Beguinage in Leuven is particularly charming. It was founded in the 13th century but with most present buildings from the 17th century. It has lovely cobbled streets and courtyards and red brick buildings deviating quite far from rectilinear. The whole site was purchased by the university, KU Leuven, in the 1960s.

Species Distribution Modelling: GeoLifeClef

This is the main thing I’ve been working on for the last ~1.5 weeks. The idea is to see how well TESSERA does at species distribution modelling (SDM). SDM is a fun idea: given a set of species occurrence records and some environmental covariates / remotely sensed images, predict the distribution of the species across a landscape. For years I’ve been interested in SDM, because it combines basically all of my interests: nature, maps, grungy Bayesian statistics to deal with the extreme observational bias.¹ I’ve dabbled a little in the past with small projects, but this is my first time really pursuing a project full time in a professional capacity. I’m excited to be working on this!

My idea for a first project here was to apply Tessera to the 2024 GeoLifeClef competition. This is an annual SDM competition hosted on Kaggle, run by the Pl@ntNet team. The idea is to make multi-species SDMs using the GeoPlant data: ~10k plant species, ~90000 presence-absence points, and ~5 million presence-only points.²

Experiment: GeoLifeClef provides a “baseline” model using just Sentinel-2 images (128x128 10m pixels, R/G/B/NIR). It ignores the presence-only data in the GeoPlant dataset, and trains only on the presence-absence data. It’s a relatively simple model. Not to say light, it’s a relatively big neural net, but just plug+play, nothing fancy at all with hyperparameter tuning / ensemble learning etc (which the winning submissions all do). Basically I took that baseline and swapped in Tessera “images”. I’m using Tessera v1.0 because 1.1 is not yet available over the whole region (which spans basically all of Europe).

Headline result: Tessera beats the baseline model by a wide margin. It gives an even better result than the baseline model with S2 supplemented with landsat+bioclim time series.

Detailed scores: These are the “private” F1 scores used to rank submissions for the GeoLifeClef competition.
1. Baseline S2-only model: 0.23629 (would have ranked 38th in competition)
2. Baseline S2 +Landsat time series + Bioclim time series model: 0.31626 (would have ranked 25th in competition)
3. Tessera-only model (i.e., model 1 with S2 images replaced by Tessera images): 0.33032 (would have ranked 11th in competition)

For reference, the top 3 scores were: 0.40890, 0.36837, 0.35292. But, as I mentioned above, these all do much more sophisticated things. Not necessarily heftier architectures, but a lot of hyperparameter tuning, ensemble stuff, and they incorporate the GeoPlant presence-only data, which the “baseline” models ignore.³

Trees of the Cambridge University Botanic Garden

I’ve visited the Cambridge University Botanic Garden many times. Last weekend was my first time really appreciating its wonderful collection of trees. My wife and I diligently followed the Trees of the Botanic Garden Trail,⁹ which took us to 13 excellent trees around the garden. Two highlights:

1. The Osage orange (Maclura pomifera). This is a tree with very odd-looking fruit, described in the trail booklet as “pickled gardeners’ brains”.

2. The Indian horse-chestnut (Aesculus indica). This is a glorious tree, and was in flower when we visited. Interestingly, the flowers all have either pink or yellow centres: yellow at first, then pink once they have been pollinated!

James Hutton Memorial Garden

Continuing the subject of visiting gardens and diligently following paper guides. My wife and I have been working through a book called “Secret Edinburgh: An Unusual Guide”. This is a guidebook full of weird and hidden spots around the city. Earlier this week it took us to the James Hutton Memorial Garden, not far from Holyrood. Despite its central location, it was extremely difficult to find. One reaches it by going into the car park of the University gym (the Pleasance Complex), then looking for the entrance to a path in between parking bays 20 and 21. This path then snakes down a steep hill to the garden.

The garden was built on the site where once stood the house of James Hutton, the 18^th century Scottish “Father of Modern Geology”.¹⁰ An interesting feature here is the collection of rocks placed around the garden. One interesting rock, just visible at the top-left of my photo above, was taken from Glen Tilt in the Cairngorms. This was the site where James Hutton observed veins of granite running through schist and limestone (as in the photographed rock). From this, he developed the theory of plutonism, that igneous rocks formed from volcanic activity, and not from crystallisation in the oceans.

Temperate Rainforests

Inspired by Guy Shrubsole’s excellent book¹ on the subject, I thought I would make a map of Britain’s temperate rainforest by training a model with TESSERA.

Some context:

• Temperate rainforests are oak-dominated or hazel-dominated woodlands that are very rainy.²
• They occur along the soggy western edge of Britain (hence: “Celtic rainforest”).
• They are very biodiverse, hosting internationally important assemblages of bryophytes/lichens/fungi etc.
• In Britain, they are rare/fragmented/degraded/overgrazed
• They are not well mapped. To my knowledge, the best map to date is the crowd-sourced one produced by Guy Shrubsole. This was a fantastic project, but due to its crowd-sourced nature is certain to be incomplete.

Given that last point in particular, I wondered: can we map Britain’s temperate rainforests from space?

In 1.5 days at the end of last week, I had an extremely quick and dirty first pass at this.³ Result below.

The map certainly seems to catch a great deal of rainforest down the western seaboard, but also tends to overpredict a little: it flags a lot of wet, woody places that aren’t rainforest, such as birch-dominated boggy woodlands in the North York Moors, or alder-dominated boggy woodland in the New Forest. In ML terms, it seems by eye that the recall is good but the precision is bad.

In the coming weeks, I’ll have a go at improving this map. Primarily, I’ll improve my training data and switch from a PU approach by getting some ‘negative’ examples. That said, we aren’t sure yet whether there is a valuable ‘science’ case (read: publication) for the map, it might just be a fun story to engage the public. If so, still worth pursuing!

Borders Abbeys Way

My wife and I have been walking the Borders Abbeys Way in stages. Last week we walked the ~22km stretch from Jedburgh to Hawick.⁵ This was rather nice, some excellent views from Black Law, and then a lovely stretch along the Teviot up to Hawick. We saw yellowhammers, dippers, and a mating pair of yellow wagtails. This was our first stretch of the Borders Abbeys Way with no (visible) goosanders.⁶

Below is a map of the route. I’ve also uploaded it as a route on the Ordnance Survey website/app.

Research

I made incremental progress on various projects. Don’t want to plant too many flags here yet until these are a little more mature. However, a couple of exciting bits:

KBAs. I’m going to be working with BirdLife International on finding new candidates for Key Biodiversity Areas (KBAs). These are areas around the world flagged as key for conservation.

The criteria for identifying KBAs are a little complicated, and the existing method for longlisting KBAs involves many repeated iterations of manual validation (and is therefore quite expensive). We’re hoping to find a clever way, likely with TESSERA (see below), to automate a little bit of the longlisting.

This past week, BirdLife shared some of their data with me and I’ve started getting my teeth into it.

TESSERA v1.1. I wrote about TESSERA in last week’s notes. This past week, a new version (1.1) has been released.¹ This addresses some issues with v1.0, such as tiling artefacts. It also goes a little farther out from the coast everywhere. I’ve spent a bit of time playing around with the new embeddings, which are very pretty!

Python package: datahues

Yesterday I tidied up and released a python package I made some months back. It’s called datahues, and has a very simple functionality: given two colours, generate a perceptually uniform colour ramp between them. I find myself needing² to do this all the time and in the past have always just bodged around with some quite complicated colour-space libraries, so I thought I would write a simple utility for it. It’s now available on PyPI and conda-forge.

Here is the figure from the README, demonstrating the need for the utility:

In the left-hand panel, the green transitions very unevenly to red: there is some awful banding in the transition zone. Without perceptual uniformity (as in the right-hand panel), one might well see patterns in data that aren’t truly there.

Here is a map I made with datahues a few months ago, depicting the annual number of sunshine hours across GB:

It is very gloomy in the Highlands!

Garden

Our vegetable seedlings had a bit of an adventure in May. They travelled 800 miles and lived in four different houses. Now, they’re finally in the ground in our garden in Edinburgh.

These are sweetcorn and yellow courgettes. We’re growing them in the same bed (along with some upcoming borlotti beans), following the “three sisters” method of companion planting. The idea is that the corn provides a rigid structure for the beans to climb, while the beans fix nitrogen in the soil. The courgettes, meanwhile, provide ground cover to retain moisture and suppress weeds.

All of our seeds were F1 varieties, a decision I now regret a little. Having recently read Dan Saladino’s excellent book Eating to Extinction, I’m a zealous convert to the idea of preserving traditional, open-pollinated seed varieties. Doing so helps preserve food heritage, rare flavours, and genetic diversity. Next year!⁴

Clerihews

I learned this week about a very silly poetic form: the clerihew. 4 lines, AABB, whimsical and funny, with the rhymes often extremely contrived and the metre absolutely all over the place. They were invented around the start of the twentieth century by Edmund Clerihew Bentley, and they are typically biographical, with the subject of the poem being the first line. For example, from Bentley’s 1905 Biography for Beginners:⁵

Sir Humphry Davy
Abominated gravy.
He lived in the odium
Of having discovered sodium.

Another example, also from Bentley’s collection:

The people of Spain think Cervantes
Equal to half-a-dozen Dantes:
An opinion resented most bitterly
By the people of Italy.

They aren’t always biographical. My favourite is from the introduction of Bentley’s collection, and is about the very nature of biography:

The Art of Biography
Is different from Geography.
Geography is about Maps,
But Biography is about Chaps.

TESSERA

TESSERA is a geospatial foundation model developed here in Cambridge. It providing a vast, rich dataset of “embeddings”: summary vectors for every 10m pixel on the planet. My new job is to do a bunch of conservation science with these embeddings.

This past week, I had my first proper play with the embeddings. I used them to train a very basic ML model, making some toy habitat maps of the Centre for Landscape Regeneration project areas (the Cairngorms, the Fens, and Cumbria). These were just quick and dirty first passes to get a ‘feel’ for Tessera.

Research plan

I spent a lot of the week thinking about my longer term research goals, both for the coming years in Cambridge and the years thereafter. I’ve written up a first draft of a research plan, just 3 pages. I won’t go into too much detail about it here, but it has been fun reading around a lot of the literature (see below!). One detail I will add here is that I used the typeface Bembo, which I’ve gotten very into recently.

100 Maps

I used the Bank Holiday Monday to get a bit of work done on my side project (“100 Maps of Britain”, hoping to have a page about that on this site soon!). One of the maps is of linguistic roots of British place names. Here, I learned about the Landsker line, a surprisingly sharp linguistic boundary in south Pembrokeshire between regions of predominantly English names (south) and Welsh place names (north). The line is thought to have been formed in the 12th century, when the Normans established themselves in the area and encouraged English settlement.

Research Literature

I read a bunch of stuff around: agri-environment schemes, causal inference, and some of the older back-catalogue of papers around the “land sharing” versus “land sparing” debate.¹

The idea here is that there are two broad approaches one could take to grow a fixed amount of food across a landscape: modern, intensive agriculture at a few sites and pristine nature reserves elsewhere (land sparing), or low-intensity, nature-friendly agriculture everywhere (land sharing).

Of course, the reality is not so much a binary as a spectrum. Nonetheless, evidence is beginning to accumulate in favour of approaches towards the land-sparing end of the spectrum.

This is a nuanced question, and I won’t be able to do it justice yet. Planning to read more about it in coming weeks.

Elsewhere

Less work-related, I’ve been reading “What the Railways Did For Us” by Stuart Hylton. A very well-researched and funny book about the coming of the railways to Britain and their impact on society, such as the impact on the economy, on cities, and even on the class system.

The opening chapter, concerning everyone’s grumbles about the initial spread of railways, was particularly memorable. Here is Wilhelm I, Kaiser of Prussia, on the subject:

Parts of the discussion struck a painfully familiar note: