On this month’s episode of 5 Minutes With, Marissa Lo (Assistant Editor) speaks to Sam Bancroft, a PhD student in the SENSE Earth Observation CDT, University of Leeds.
Sam Bancroft, a PhD student in the SENSE Earth Observation CDT, University of Leeds (Image credit: Marissa Lo).
[00:10] Marissa Lo: Hello, and welcome to 5 Minutes With. My name is Marissa Lo, and today I’m joined by Sam Bancroft, a PhD student at the University of Leeds. Thanks so much for joining us, Sam. Can you tell us about your PhD research?
[00:24] Sam Bancroft: My PhD is to do with assessing future food production, assessing food security, particularly in areas and countries where this is going to be an ever-pressing question over the next few decades. The way I do that is that I look at crops from space. I look at lots of satellite imagery that streams down to the Earth every day, and I use different machine learning techniques to distinguish crop types from that. So, I can look at the way that crops are growing throughout the calendar year, and I can say, that’s maize, that’s wheat, that’s linseed. And I can do this in a way that I don’t need to hard code or explicitly tell a computer how to work out what to do. I employ machine learning techniques, which is I train a model, I teach it with a few labelled examples, and then I, hopefully from that, have a successfully trained model that I can set free into the wild. And it works to detect crops over new countries and new periods of time. What I’m doing with the crop maps that I make from these models is I’m making sure that they’re generalisable. And what I mean by that is that they operate effectively and accurately against data that it hasn’t seen before. And I’m making it more generalisable in space and time so I can deploy it for, like, national mapping projects in a way that currently present day can’t be achieved very easily. And once I have a national scale crop map, I look at machine learning techniques again to look at the causal mechanisms for agricultural dynamics. So I look at why farmers are following certain pathways to develop their crops and what kind of picture we can be expecting in the next 10, 20 years. Is there going to be a shift away from peasant farming to corporate farming? Are cash crops going to be favoured over more consistent long term cereal crops? How is this going to be affecting people’s nutrition? How is this going to be affecting people’s finances, the soil health, and so on. So, there’s many different directions you can go with that.
[02:08] Marissa Lo: So, what’s your favourite thing about your project and your research area?
[02:12] Sam Bancroft: I think the favourite thing about my project it’s a very pressing issue that straddles both industry and research. So, there’s a lot of motivated and exciting people out there working on quite similar problems, but there’s enough interesting issues and questions to go around for everyone. It’s not something theoretical that is resigned to the dark corridors of academia. Companies can actually offer this commercially, and they actually offer a product in which people are willing to pay a lot of money to work out how their farms are operating. Employing the similar techniques that I’m doing in academia, which is hopefully addressing some of the problems caused by climate change. But also food underpins a lot of the structures of society, it’s quite fundamental to who we are, and it kind of reaches into many different pockets and the kind of work that I do. So while I’m just working on crops, I get to dip my toes into so many different exciting areas at the same time.
[02:57] Marissa Lo: So finally, what advice would you have to readers of Geoscientist who want to learn more about machine learning or get involved in this research area?
[03:06] Sam Bancroft: So what I do, I, very much, I separate that into two categories: machine learning and remote sensing. Very much for machine learning I recommend, for those who want to get into it, try by doing. It’s a very hyped-up field and while, increasingly, courses might be offered at universities, machine learning is quite a tricky field to get to grips with. Sometimes I would recommend websites like Omdena and Kaggle, where you can do machine learning competitions. A famous teaching tutorial example is that everybody starts looking at the Titanic data set, which is a data set where you can create a simple machine learning model to predict: would you make it? Would you survive the Titanic sinking based on your social status or your room number, et cetera? That’s an exercise that people can do to teach themselves the concepts, and it’s much better than sitting in a lecture or watching a video about it, in my opinion. And then for remote sensing, I very much got into remote sensing because I did a physics degree with astronomy and I realised that all these astronomy telescopes, I can flip them the other way around and look down at the Earth instead, for Earth observation. I could recommend, if perhaps you’re not from a geographical or physics or physical science oriented degree, that you look at the Sentinel satellites from the European Space Agency, mainly because that’s open access data, quite easy to get to grips with, and you can look at it online in your browser without having to download anything. You can very much get started with what’s free and openly accessible.
[04:27] Marissa Lo: Thanks very much, Sam, and best of luck with the rest of your project.
[04:30] Sam Bancroft: Thank you very much. Thanks for having me.