Starting my PhD on the mapping of flooding
Specifically, using new technologies such as AI and the Internet of Things to map and predict where it's going to flood in real-time.
This year, I'm starting a 3 year funded PhD on dynamic flood risk mapping, as part of a cluster of water-related PhDs that are all being funded at the same time. I've got some initial ideas as to the direction I'm going to take it too, which I'd like to talk a little bit about in this post.
I've got no shortage of leads as to potential data sources I can explore to achieve this. Just some of the avenues I'm exploring include:
- Monitoring street drains
- Fitting local council vehicles with sensors
- Analysing geotagged tweets with natural language processing techniques
- Predicting rainfall with aggregate mobile phone signal strength information
- Vehicle windscreen wipers
- Setting up static sensors? Not sure on this one.
Also, I've talked to a few people and thought of some pre-existing data sources that might come in useful:
- Elevation maps
- Vegetation maps
- Normalised difference vegetation index - Map of vegetation density that's already available
- Normalised difference water index - detects water on the surface of the earth - including that contained within leaves
- River levels
Finally, I've been thinking about what I'm going to be doing with all this data I'd potentially be collecting. First and foremost, I'm going to experiment with InfluxDB as a storage mechanism. Apparently it's supposed to be able to handle high read and write loads, so it sounds suitable a first glance.
Next, I'm probably going to wind up training an AI - possibly incrementally - to predict flooding. Unlike my summer project, I'm probably going to be using a more cutting-edge and exotic AI architecture.
I suspect I might end up with a multi-level system too - whereby I pre-analyse the incoming data and process it into a format that the AI will take. For example, if I end up using geotagged social media posts, those will very likely filter through an AI of some description that does the natural language processing first - the output of which will be (part of) the input (or training output?) for the next AI in the chain.
I've given some thought to training data too. My current thinking is that while some data sources might be used as inputs to a network of interconnected AIs, others are more likely to take on a corrective role - i.e. improving the accuracy of the AI and correcting the model to fit a situation as it unfolds.
All this will not only require huge amounts of data, but will also require a sophisticated system that's capable of training an AI on past datasets as if they were happening in real-time. I suppose in a way the training process is a bit like chronological history lessons at speed - catching the AI up to the present day so that it can predict flood risk in real-time.
Before all this though, my first and foremost task will be to analyse what people have done already to map and predict flood risk. Only by understanding the current state of things will I be able to improve on what's already out there.
Found this interesting? Got any tips? Comment below!