Modelling the range of wild plants and crop progenitors in the Late Epipalaeolithic

Gronenborn et al. 2021, https://doi.org/10.5281/zenodo.5903165

Hi, my name is Joe Roe, I work at the Institute of Archaeological Sciences here in Bern. Today I’d like to talk about computational models of the geographical range of plants in past environments.

My background is in computational archaeology, broadly speaking, but also in West Asian prehistory. So I’m interested in using these models to understand the ecological context of the emergence of agriculture in the Neolithic of West Asia. This represented, as far as we know, the first settled farming societies in the world, as well as the centre of origin for later agricultural economies across Europe, North Africa and a large part of Asia. So although there were of course other places were domestication took place independently, the West Asian Neolithic is really chapter one in the story of the prehistory of farming, in most of the world.

When we talk about the emergence of agriculture in those regions were domestication took place (as opposed to where already-domesticated crops were transplanted), there’s no clean break between “forager” and “farmer”, or “wild” and “domestic”. Instead we have to recognise a sequence of stages in people’s relationship with (in this case) plants. Domestication—the alteration of an organism’s physiology due to human selection—is the clearest sign of the emergence of an agricultural economy, but by definition it must be preceded by a phase of “pre-domestication cultivation”, which in turn could well be preceded by forms of management that sit somewhere between ‘gathering’ and ‘cultivation’. Conversely, the appearance of domestic crops does not mean that the society is now a farming one; ‘cultivation’ can be a relatively low-intensity activity (think scattering seeds into a floodplain and coming back a few months later), and can still represent a minor part of the subsistence base in comparison to wild foods.

Plants have short generations, so theoretically this whole process (cultivation to domestication to full-blown agriculture) could happen in a matter of decades. At one point this was indeed what we thought happened in West Asia – there was a small region, a “golden triangle”, where people began cultivating, and then rapidly domesticated, a core package of “founder crops” right at the beginning of the Neolithic.

Arranz-Otaegui et al. 2016, https://doi.org/10.1073/pnas.1612797113

Arranz-Otaegui et al. 2018, https://doi.org/10.1016/j.quascirev.2018.02.011

Arranz-Otaegui & Roe 2023, https://doi.org/10.1007/s00334-023-00917-1

Today, I think it’s fair to say that the consensus has shifted to recognising a much more protracted transition. We can see the earliest signs of pre-domestication cultivation well into the Epipalaeolithic, perhaps as far back as 20,000 years ago during the height of the last glaciation. We also now know, from genetic and physiological studies, that the domestication of plants generally happened very slowly, so slowly in fact that it was probably imperceptible to the humans involved. And post-domestication, we can see that it took a long time, and a lot of experimentation with many different plants, before anything like the familiar agricultural economy based on arable crops emerged – probably not until well into the latter part of the Pre-Pottery Neolithic, after c. 8000 BC.

With this stretching of the timeline we’ve also, probably unsurprisingly, also come to recognise that the transition took place over a very wide area (about the same size of Europe) with a very diverse range of ecologies and domesticable species available. And this was a time of significant climatic change—the Pleistocene to Holocene transition—so that map will have been constantly changing.

For me, this radical recent change in our understanding of the pace and extent of the transition to agriculture really emphasises the importance of understanding its biogeographical context. Of course, we have archaeobotany and palynology, and so on, that gives us a wealth of information on the flora of particular times and places. But what’s frequently missing is an idea of what was growing in the wider region, and how that changed through time:

What was the “wild range” of crop progenitors?
When did they start appearing outside it?
Do the plants we observe on archaeological sites merely reflect what was available in the local environment?

I think these questions are really key to a fuller understanding of Neolithic economies.

To try and get a clearer picture of this biogeographical/ecological context, I use an approach that’s widely used in ecology, called species distribution modelling or, in this case more correctly, ecological niche modelling (ENM). I think the best way to illustrate what that is is to show you what kind of results you can get with it. So I’ll start with that – And explain more about how it actually works at the end of the presentation.

Flax

First, a nice straightforward example. These maps show the modelled distributions of flax.

The snapshots correspond to aggregated climatic periods, from the Last Glacial Maximum through to the Early Holocene (roughly coinciding with the Pre-Pottery Neolithic), to the modern day (the training data). The date ranges given are in kiloannum (thousand years ago) notation.

The prediction is probabilistic. Not the probability of occurrence exactly; but the probability of classifying a location as “present”, according to the model. In loose terms: how certain the model is about predicting occurrence in each cell. The lilac colour indicates zero, green is mid-values, and blue–purple is high certainty.

Flax has traditionally been considered one of the “founder crops” of the Neolithic, but according to our analysis, it is actually barely present in Neolithic sites. Here we can see that the natural range is very small and is not predicted to have changed very much over the last 20,000 years ago So this perhaps explains its limited distribution in the earlier Neolithic; it is simply a plant with a limited range, and probably needed more time to adapt to wider environements than other, more naturally widespread crops. A modest finding – but what I think is important is that without this biogeographic background, we might be tempted to interpret the low ubiquity of flax on archaeological sites as an active choice not to use it, when in fact it looks just like .

Rye

A perhaps more interesting result is for rye.

Rye has always been a bit of a puzzle, because it is native to West Asia, and it occasionally show up in Neolithic archaeobotanical assemblages there. But it doesn’t show up in large numbers, its domestic form, until much later, in the European Neolithic – so the question is where did it come from? Is it a West Asian domesticate, in which case why is it so rare there? Or did it perhaps spread with other cereal crops as a weed, then undergo a secondary domestication event in Europe?

The model predicts that the prehistoric distribution of rye was significantly smaller than the modern range, notably excluding much of central and eastern Turkey, which were otherwise hotspots for domestication. Tentatively, I think this speaks against the weed-origin hypothesis: with this restricted range, it seems implausible that rye would have spread without more active human intervention. On the other hand, the indicated range suggests that we might find the origin of domestic rye in some unusual places: the Southern Levant, maybe, but with stronger confidence on Cyprus (not usually considered part of the centre of origin), western Anatolia, or even Greece. This is all very tentative – but interesting!

Bread wheat

Polyploid wheat origins (Ian Alexander, redrawn after Golovnina et al. 2007, https://doi.org/10.1007/s00606-006-0478-x)

Bread wheat

Benjamin Nowak, https://twitter.com/BjnNowak/status/1648954308170338305

Taxon	Occurrences (Neolithic)	Occurrences (GBIF)
Triticum turgidum dicoccum (incl. Triticum aestivum, Triticum dicoccoides, Triticum dicoccum)	64 (44%)	62
Hordeum spontaneum (incl. Hordeum spontaneum, Hordeum vulgare)	62 (43%)	6550
Triticum monococcum aegilopoides (incl. Triticum boeoticum, Triticum monococcum)	44 (30%)	2902
Bolboschoenus maritimus (incl. Bolboschoenus maritimus, Scirpus maritimus)	33 (23%)	281
Vicia ervilia	33 (23%)	672
Buglossoides tenuiflora	31 (21%)	134
Arnebia decumbens	25 (17%)	227
Buglossoides arvensis	25 (17%)	122
Medicago radiata	21 (14%)	873
Androsace maxima	20 (14%)	68
Vicia orientalis (incl. Lens culinaris)	17 (12%)	367
Medicago astroites (incl. Trigonella astroites)	16 (11%)	118
Arnebia linearifolia	14 (10%)	148
Gypsophila vaccaria (incl. Vaccaria pyramidata)	11 (8%)	292
Linum bienne (incl. Linum bienne, Linum usitatissimum)	11 (8%)	100
Carex divisa	10 (7%)	156
Ficus carica	10 (7%)	2463
Lathyrus oleraceus (incl. Pisum elatius, Pisum sativum)	8 (6%)	1380
Vicia faba	8 (6%)	1734
Aizoanthemopsis hispanica (incl. Aizoon hispanicum)	7 (5%)	288
Pistacia atlantica	7 (5%)	725
Prosopis farcta	7 (5%)	1084
Rumex pulcher	7 (5%)	621
Ammi majus	6 (4%)	408
Cicer reticulatum (incl. Cicer arietinum)	6 (4%)	193
Hordeum bulbosum	6 (4%)	1755
Salsola kali	6 (4%)	267
Taeniatherum caput-medusae	6 (4%)	143
Capparis spinosa	5 (3%)	3180
Chenopodium album	5 (3%)	218
Lolium temulentum	5 (3%)	86
Poa bulbosa	5 (3%)	1144
Aegilops crassa	4 (3%)	143
Atriplex prostrata	4 (3%)	53
Brachypodium distachyon	4 (3%)	1281
Fumaria densiflora	4 (3%)	143
Gypsophila pilosa	4 (3%)	67
Hordeum murinum	4 (3%)	914
Lathyrus aphaca	4 (3%)	1657
Lathyrus sativus	4 (3%)	241
Secale cereale	4 (3%)	522
Triticum durum	4 (3%)	167
Avena sterilis	3 (2%)	47512
Bassia arabica	3 (2%)	228
Bromus sterilis	3 (2%)	254
Cephalaria syriaca	3 (2%)	135
Helianthemum ledifolium	3 (2%)	628
Lolium rigidum	3 (2%)	1194
Phalaris paradoxa	3 (2%)	298
Quercus ithaburensis	3 (2%)	1154
Triticum aestivum compactum	3 (2%)	122
Verbena officinalis	3 (2%)	853
Vicia narbonensis	3 (2%)	1297
Aegilops speltoides (incl. NA)	NA	1379
Aegilops tauschii (incl. NA)	NA	666
Triticum urartu (incl. NA)	NA	2005

So where do these models come from? I could of course spend a long time talking about the choice of modelling approach, parameterisation, data cleaning, and so on—and I’m very happy to, if anyone wants to find me afterwards!—but conceptually the method is really very simple. All you really need to know to understand where these predictions come from is this.

The model is trained on two sets of data:

Species occurrences – that is, the coordinates of where the target species has been observed growing in the wild. Most people get these from an excellent XXXX
Environmental variables, typically climatic ones (precipitation, temperature, etc.), but you can also throw in topography, soil characteristics, anything you can get that you think is relevant to where plants grow, basically.

Importantly, I use contemporary occurrence data and contemporary environmental data. This is different to how some archaeologists have used ENM in the past, which is to use archaeological occurrence data – from archaeobotany, zooarchaeology, and so on. But personally I think it makes more sense to start with the high quality data we have on modern plant biogeography, to train the best model possible, and keep the archaeological record in reserve as a verification dataset.

To train our model from these datasets, there are a number of options. I use a machine learning method called Random Forest, but there are also more standard regression approaches, Bayesian approaches, another machine learning technique called MaxEnt is also popular… But whichever you use, what you get out of it is a model that can be used to predict the occurrence of a species, given a set of environmental variables.

So the final missing part of the puzzle is some paleoclimate data, which we can use to ‘turn back the clock’ on those environmental variables. Luckily, it’s getting easier and easier to access hindcasted outputs from global climate models – those gigantic models the IPCC use to predict future climate change, with the dial turned backwards. I’ve for a long time used a great dataset called PaleoClim, which has global, high resolution data for the whole world, aggregated into the kind of climate period slices I’ve shown here. But I’m thinking for the next stage of this research I’ll switch to continuous predictions, since those are now quite easy to get hold of.

Franklin et al. 2015, https://doi.org/10.1016/j.quascirev.2014.12.015

Modellers ❤️ Open Data

https://gbif.org

http://www.paleoclim.org/

(R 📦: https://github.com/joeroe/rpaleoclim)

I also want to highlight that the methodology I’ve presented here is entirely dependent on open data. GBIF, PaleoClim, and similar datasets are extraordinarily useful resources for computational modelling in archaeology. And since I’ve been playing around with this approach for some time, I can say that year on year the release of new datasets (and tools) keeps making it easier and easier to apply.

That does mean, unfortunately, that archaeology is the weak link. Getting together archaeobotanical and zooarchaeological data to verify these models, as I’ve suggested could be very useful here, is still not easy. That’s why I’m really glad and grateful to see so much emphasis on collating, systemising and openly publishing different types of archaeological data. For people like me, whose research relies almost entirely on data reuse, it’s essential.

Modelling the range of wild plants and crop progenitors in the Late Epipalaeolithic–Early Neolithic Levant

Joe Roe
University of Bern

Contact

https://joeroe.io
joeroe@hey.com
joeroe@archaeo.social

These slides

joeroe.io/dabawa23_enm
joeroe/dabawa23_enm
10.5281/zenodo.8416267.

Key references: Arranz-Otaegui, A. & Roe, J. 2023. Revisiting the concept of the ‘Neolithic Founder Crops’ in southwest Asia. Vegetation History and Archaeobotany. https://doi.org/10.1007/s00334-023-00917-1; Brown et al. 2018. PaleoClim, high spatial resolution paleoclimate surfaces for global land areas. Nature Scientific Data. https://doi.org/10.1038/sdata.2018.254; Franklin et al. 2015. Paleodistribution modeling in archaeology and paleoanthropology. Quaternary Science Reviews. https://doi.org/10.1016/j.quascirev.2014.12.015; Valavi et al. 2021. Modelling species presence-only data with random forests. Ecography. https://doi.org/10.1111/ecog.05615
Acknowledgements: Amaia Arranz-Otaegui, Peter Yaworsky, Tobias Richter, Martin Hinz, Albert Hafner; C. L. Davids Fond og Samling, Swiss National Science Foundation